论文解读
AI 辅助整理- 方法
- 构建165类73.3万图像数据集,渐进训练蒸馏基础模型先验。
- 发现
- 零样本掩膜AP50:95达69.06,超SAM 3 4.76,可边缘部署。
- 机会点
- 可探索蒸馏分割器在采摘机器人实时感知与多模态边缘部署的泛化边界。
论文摘要
提出水果实例分割数据构建与渐进训练框架SAF,把基础模型先验蒸馏为轻量无提示分割器。覆盖165类水果、73.3万图像、966万整果掩膜;零样本SAF-RF-DETR-XL掩膜AP50:95达69.06,较SAM 3高4.76,可边缘部署。
全文
arXivSegAnyFruit: Learning Generalizable and Edge Deployable
Fruit Instance Segmentation at Scale
Yanan Wang1 Yuzheng Wang2 Yibin Ying1 Zhenghao Fei3*
1Zhejiang University 2Alibaba Token Hub, Alibaba Group 3University of California, Davis
{mmwang@zju.edu.cn, zfei@ucdavis.edu}
Figure 1.Overview of SegAnyFruit.(A) Three-stagePre–Mid–Posttraining recipe. (B) Prompt-free fruit segmentation across diverse
scenes and unseen types.. (C) Efficient automatic annotation. (D) Cross-domain recipe validation. (E) Scale of our curated resources.
Abstract
Deploying fruit instance segmentation in real agricultural
settings requires generalizing across unseen species, or-
chards, and imaging conditions under strict onboard com-
pute and annotation budgets. We introduceSegAnyFruit
(SAF), a data construction and progressive training frame-
work that distills foundation-model priors into lightweight,
prompt-free segmenters. SAF combines generic mask pre-
training, broad mid-training on automatically annotated
web imagery from our curatedWebFruitsdataset, and
human-verified agricultural refinement post-training onPa-
perFruits, together spanning 733K images and 9.66M
whole-fruit masks across 165 fruit types. Across seven
benchmarks, zero-shotSAF-RF-DETR-XLreaches 69.06
average mask AP50:95 (+4.76 over SAM 3) and deployable
on edge device. SAF initialization also boosts few-shot to
full-shot adaptation across five agricultural domains while
*Corresponding author: Zhenghao Fei, Department of Biological and
Agricultural Engineering, University of California, Davis, CA, United
States.
mitigating catastrophic forgetting. Finally, we systemati-
cally analyze training stages, data scale, model capacity,
and generative augmentation, and validate our pipeline’s
transferability to underwater, cell and human segmentation.
We will release our datasets, code, and SAF-models.
- Introduction
Reliable fruit instance segmentation is critical for robotic
harvesting, yield estimation, growth monitoring, and qual-
ity inspection. However, achieving robust instance segmen-
tation across diverse visual environments with limited an-
notations remains a central challenge for agricultural per-
ception. Fruits exhibit substantial variation in shape, color,
size, and texture across varieties, species, growth stages,
and cultivation practices, while dense foliage, occlusion,
and changing illumination further complicate localization
and instance separation [15, 26].
Closed-set specialist models, including YOLO-based ar-
chitectures, have achieved strong in-domain accuracy on in-
1
arXiv:2611.25475v1 [cs.CV] 19 Nov 2026
dividual fruit benchmarks [6, 22, 30, 63]. However, gener-
alization to unseen orchards and unfamiliar fruit types re-
mains challenging as acquisition conditions and visual dis-
tributions change [3, 31]. Deploying these specialists in
new settings therefore often requires renewed data collec-
tion, manual annotation, and fine-tuning.
Foundation models offer a promising alternative to
dataset-specific training. Large vision-language models
(VLMs), such as Qwen3-VL [2], and open-vocabulary
detectors including OWLv2 [38], Grounding DINO [35],
DINO-X [49], YOLO-World [9], and YOLOE [57] ex-
tend recognition beyond fixed categories, while the Seg-
ment Anything family provides promptable instance seg-
mentation [7, 28]. Recent fruit-focused efforts also
build on foundation-model knowledge: FMFruit focuses
on bounding-box detection [31], and SDM-D distills
foundation-model supervision into compact students, us-
ing fruit-specific appearance descriptions during annota-
tion [58]. The computational demands, inference latency,
and potential API dependence of frontier VLMs and large
promptable segmenters further motivate efficient local in-
ference on agricultural robots operating in the field.
In this work, we argue that achieving foundation-model-
level generalization in a specialized perception task does not
require a massive, general-purpose model. Instead, transfer-
ring foundation-model training principles, large-scale pre-
training, diverse supervision, and task-specific refinement,
can enable a compact segmenter to learnfruitas a uni-
fied, generalizable visual concept across species and envi-
ronments without target-specific prompts at inference.
To this end, we introduceSegAnyFruit(SAF), a data
construction and progressive training framework for gen-
eralizable fruit instance segmentation (Figure 1). Starting
from public pretrained weights, our SAF Training Recipe
pretrains SAF models on large-scale general segmentation
dataset SA-1B [28] to acquire generic mask priors, mid-
trains them on automatically annotated WebFruits to learn
diverse fruit appearances and boundary cues, and post-
trains them on human-calibrated PaperFruits from dense
orchard environments to refine agricultural segmentation.
The SAF model family offers accuracy–speed trade-offs, in-
cluding compact variants for edge deployment. Our SAF-
RF-DETR-XL surpasses the evaluated VLMs and open-
vocabulary baselines in average mask AP50:95 across seven
benchmarks. SAF initialization supports label-efficient
adaptation and anti-forgetting across five unseen agricul-
tural domains. Our contributions are fourfold:
•A progressive, cross-domain training recipe.We pro-
pose a three-stage training recipe (Pre on large-scale
general segmentation data, Mid on curated auto-labels,
and Post on human-optimized annotations) that distills
foundation-model priors into lightweight, prompt-free
segmenters. Primarily validated on fruit perception, this
training recipe also generalizes effectively to underwater,
cell and human segmentation under scarce labels.
•Large-scale, instance-curated fruit datasets.We con-
struct and will release WebFruits, PaperFruits, and seven
manually annotated evaluation benchmarks, spanning
165 fruit types, 732,870 images, and 9.66M masks, paired
with a category-aware curation pipeline enforcing whole-
fruit instance granularity.
•An open-source model family for zero-shot, adaptive,
and edge deployment.We release edge-friendly SAF-
models across multiple architectures and scales. Zero-
shot SAF-RF-DETR-XL reaches 69.06 average mask
AP50:95 and SAF-YOLO26m-seg reaches 61.34, while
their FP16 TensorRT engines run at 30 FPS and 69 FPS
on Jetson Orin NX, respectively. Beyond zero-shot in-
ference, SAF initialization improves few-shot to full-shot
adaptation while preserving non-target performance.
•Systematic study of supervision, scale, and label effi-
ciency.We analyze how training stages, data scale, model
capacity, and generative video augmentation govern gen-
eralization and adaptation, offering practical guidelines
for label-efficient domain segmentation.
2. Related Work
Domain-specific visual models.Fruit perception has
evolved from hand-crafted color and geometry heuristics
to learned detectors and segmenters. DeepFruits and or-
chard Faster R-CNN [48] established deep multi-fruit de-
tection [3, 51], Polar-Net improved mask boundaries under
foliage occlusion [26], and recent YOLO variants prioritize
real-time throughput [27]. Yet strong in-domain accuracy
rarely transfers across orchards, imaging setups, or unseen
species, and fixed classification heads prevent recognition
beyond the training taxonomy [15, 35]. Rather than training
per-dataset specialists, we learnfruitas a unified visual con-
cept that generalizes across scenes and unseen fruit types at
field-ready latency.
Foundation models and open-vocabulary perception.
VLMs such as Qwen ground instructions into boxes and
points [2], Grounding DINO enables language-guided lo-
calization [35], and the Segment Anything family provides
promptable instance segmentation [7, 28, 47]. While these
frontier models generalize broadly, their heavy compute and
latency prohibit real-time deployment on onboard agricul-
tural hardware, and VLMs often lack mask boundary pre-
cision in dense foliage. Lighter open-vocabulary archi-
tectures such as YOLO-World, YOLOE, and EfficientViT-
SAM improve inference efficiency [9, 57, 64], yet remain
reliant on text prompts and underperform on small, heav-
ily occluded fruits. In agriculture, FMFruit explores open-
set multi-fruit detection but outputs bounding boxes rather
than instance masks [31], while SDM-D distills SAM 2 [47]
and OpenCLIP [10] into compact students but depends on
2
Figure 2. WebFruits and PaperFruits. (A) Coverage and distributions. (B) Image and mask counts; Train includes the internal validation
split, and PaperFruits Test summarizes the five held-out agricultural benchmarks. Collection labels include varieties and visual forms. (C)
Whole-fruit mask curation. (D) Scene diversity.
fruit-specific appearance descriptions [58]. Unlike these ap-
proaches, we target a compact, prompt-free fruit segmenter
that delivers accurate instance masks across diverse envi-
ronments while running locally on edge hardware.
Representation learning and synthetic data.Strong
visual representations reduce reliance on exhaustive man-
ual labels. DINOv3 learns transferable dense features via
large-scale self-supervision [54] for downstream segmenta-
tion. Similarly, Depth Anything and DeFM show that com-
bining broad pseudo-labeled supervision with distillation
yields compact, generalizable models for robotic percep-
tion [44, 60], extending Noisy Student’s self-training [59].
In parallel, generative augmentation offers complementary
visual coverage: DiverGen scales synthetic diversity for
instance segmentation [14], and GAN-based crop simula-
tion aids fruit-domain transfer [15]. Building on these in-
sights, we systematically examine how self-supervised and
generic mask priors, curated web pseudo-labels, human-
optimized agricultural data, and prompted video synthesis
interact across model scales and label budgets.
3. Data and Model Construction
To learn a prompt-free segmenter that generalizes across
fruit species and environments, We curate two comple-
mentary corpora, WebFruits for broad visual diversity and
PaperFruits for consolidated agricultural benchmarks with
higher-fidelity supervision (Figure 2), and train compact
instance segmentation models through a progressive three-
stage recipe (Pre,Mid, andPost)
3.1. Data Collection
WebFruits and WebOOD.Using a unified web-crawling
pipeline, we harvest∼1.68M candidate images from
Flickr [16], iNaturalist [24], GBIF [17], and web search
engines (Figure 2A(b)). Search queries combine multilin-
gual common/scientific names and varieties withcultivation
contexts(e.g., orchards, greenhouses, hydroponics, harvest-
ing) andappearance states(e.g., unripe, overripe, diseased,
cross-sectional views), such as “Malus domestica,” “apple
picking farm,” and “unripe green apple.” To ensure strict
vocabulary separation, we partition query terms into two
disjoint sets and verify separation via image metadata anal-
ysis alongside readability checks, perceptual-hash dedupli-
cation, and mask curation. This yieldsWebFruitswith
571,494 images (570,459 for training/validation and 1,035
held out asWebTest), andWebOODwith 395 manually an-
notated images spanning 38 unseen fruit types strictly ex-
cluded from training and prompt selection.
PaperFruits.We consolidate 36 agricultural and clas-
sification datasets, reserving five agricultural benchmarks
for evaluation and 31 sources for training and validation.
Representative sources include WGISD [52], MinneAp-
ple [20], ACFR [3], and VegFru [21], combining field
imagery with broad fruit appearance diversity. Paper-
Fruits comprises 160,981 images spanning orchards, green-
houses, robot-captured scenes, and classification collections
(160,156 train/val and 825 test; Figure 2D).
3.2. Data Annotation and Quality Control
WebFruits: scalable automatic annotation.We anno-
tate WebFruits using SAM 3 [7]. For each fruit collection,
3
we evaluate three prompt formats on 50 randomly sampled
images: the generic label (“fruit”), the specific category
name (e.g., “apple”), and their union (“apple or fruit”). We
then select the format that achieves the best coverage and
instance granularity for batch annotation.
Category-aware whole-fruit mask curation.Raw
SAM 3 proposals often duplicate instances, merge fruit
clusters, or over-segment surface parts. We apply a
three-step curation pipeline without altering retained mask
boundaries (Figure 2C): (1)Duplicate suppression (Overlap
NMS)retains only the higher-confidence mask when two
proposals share>90%intersection over each mask’s area;
(2)Group suppression (Geometry Filter)discards a par-
ent mask if it contains at least two retained children (each
90%inside the parent) whose union covers>60%of
the parent’s area—relaxed to one child for banana/plantain
and tomato/cherry-tomato to prevent cluster merging; and
(3)Semantic selection (VLM + checks)resolves compound
fruits such as blackberries by querying a VLM to retain
whole-fruit masks (confidence≥0.75) while rejecting indi-
vidual drupelets, clusters, and background regions. Grape,
raspberry, and mulberry undergo only duplicate suppression
to avoid discarding valid cluster units. Images yielding no
retained masks are excluded.
PaperFruits: human-optimized annotation.The 31
training sources from PaperFruits follow four annotation
routes tailored to their native labels: 3 datasets (WGISD,
Berremangra, Dyson RGB) [4, 52, 55] retain native masks;
9 box-labeled datasets (e.g., Litchi-UA V , Oranges Field,
AgRobTomato) [32, 36, 37] prompt SAM 3 with original
boxes; 18 classification/box-free sources (e.g., FruitBench,
FruitVision, VegFru) [5, 21, 56] use text-prompted SAM 3;
and Schram03 [53] uses boxes and points derived from fore-
ground priors. Six mask/box sources are further supple-
mented with text-prompted proposals. To broaden cover-
age, we merge proposals from multiple prompts, filter them
by confidence, overlap, and area limits while protecting na-
tive masks. After a final pass to remove near-duplicates
(details in the supplement), we manually inspect a sampled
subset of labeled data to prevent systematic errors.
All test images are manually annotated.
3.3. Dataset Statistics and Splits
Training corpora.As summarized in Figure 2B, the
WebFruits train/val pool comprises 570,459 images and
6,891,505 masks across 127 collection types, while Pa-
perFruits comprises 160,156 images and 2,724,216 masks
across 92 types (spanning 165 total fruit types including va-
rieties and visual forms, Figure 2A(c)). WebFruits spans
six major botanical/culinary groups, namely pome, drupes,
berries, citrus, melons/fruit vegetables, and tropical/other
fruits (Figure 2A(a)), across dining tables, markets, road-
side trees, and diverse ripeness states (Figure 2A(b), D).
PaperFruits emphasizes dense agricultural cultivation. Both
corpora span sparse and crowded scenes (Figure 2A(d)), av-
eraging 12.1 and 17.0 instances per image (with 4.9% and
8.8% of images containing≥50masks), providing strong
supervision for separating touching instances.
Evaluation benchmarks.We evaluate across seven
manually annotated test sets spanning three generalization
regimes (Figure 2B).WebTestmeasures held-out perfor-
mance within the WebFruits vocabulary, whereasWebOOD
tests zero-shot transfer to 38 unseen fruit species. Between
these regimes, five agricultural datasets, strictly excluded
from training and model selection, probe real-world do-
main shifts:MSU-Apple[11] captures clustered apples un-
der daylight foliage occlusion;KFuji RGB-DS[18] images
apples under artificial nighttime lighting to test intra-class
acquisition shifts;MangoFruitBD[25] spans healthy and
diseased mangoes;StrawDI Db1[45] features small, un-
ripe, and occluded plantation strawberries; andRob2Pheno
Tomato[1] averages 78.3 tomatoes per nighttime green-
house image to stress ultra-dense instance separation.
3.4. Model-Agnostic Progressive Training Recipe
Beyond the fruit datasets, our SAF Training Recipe em-
bodies a generalizable, architecture-agnostic paradigm for
label-efficient domain segmentation (Figure 1A). Its core
principle is to decouple supervision into three progres-
sive stages: (1)Pre-training (Pre)on generic segmentation
data builds class-agnostic objectness and boundary priors;
(2)Mid-training (Mid)leverages foundation models (VLMs
and SAM) to automatically annotate diverse in-the-wild im-
agery, replacing expensive manual labeling with scalable,
curated pseudo-supervision; and (3)Post-training (Post)re-
fines dense instance separation and domain-specific bound-
aries using a compact set of human-calibrated annotations.
In our fruit instantiation, these three stages correspond to
SA-1B [28] (Pre), WebFruits (Mid), and PaperFruits (Post).
Crucially, the recipe imposes no custom network modules
or specialized losses—training standard detectors directly
through their native objectives. We instantiate it across
three distinct architectural families: CNN-based YOLO26-
seg [27], DETR-style RF-DETR [50], and frozen DINOv3-
B/16 [54] with Mask2Former [8]. Foundation-model teach-
ers and text prompts are used only during data construction;
compact SAF models perform fast, prompt-free inference.
- Experiments
Evaluation protocol and baselines.We evaluate mod-
els on seven manually annotated benchmarks from Sec-
tion 3.3, defining OOD relative to our fruit training collec-
tions and vocabulary. Following COCO evaluation [34],
we report mask AP 50:95 and mask AP 50. We compare
SAF against general-purpose VLMs and open-vocabulary
segmenters using their strongest of three tested prompts,
4
Table 1. Fruit segmentation without target-domain adaptation. Entries: mask AP 50:95/AP50. Avg. AP 50:95 is the equal-weight average of
mask AP50:95 across the seven datasets. The six tests other than WebTest are OOD relative to our fruit training.ℓdenotes the long side.
†VLM inputs use native resolution, downscaled toℓ= 1280only whenℓ >1280. Best and second best areboldand underlined .
Model Input WebTest MSU-Apple KFuji-Apple Mango StrawDI Tomato WebOOD Avg. AP50:95
General-Purpose Vision–Language Models
GPT-6-Astra [40] Native† 41.30 / 69.19 33.22 / 63.32 25.83 / 53.06 84.74 / 96.01 58.58 / 81.39 37.84 / 72.45 49.60 / 75.69 47.30
GPT-6.1-Terra [43] Native† 31.17 / 52.73 34.57 / 62.26 31.08 / 56.84 75.05 / 89.90 55.96 / 79.68 30.69 / 54.30 35.62 / 56.46 42.02
GPT-6.1-Sol [42] Native† 38.72 / 66.93 29.32 / 59.09 25.14 / 52.93 81.90 / 94.84 57.01 / 80.16 33.08 / 66.12 45.80 / 73.18 44.42
GPT-6-Luna [41] Native† 7.14 / 16.91 8.29 / 22.69 9.21 / 24.48 57.93 / 88.45 37.51 / 67.41 4.90 / 13.43 8.63 / 19.01 19.09
DeepSeek-V4.1 [12] Native† 3.62 / 9.07 3.39 / 11.69 2.52 / 9.86 50.25 / 87.54 24.98 / 57.81 2.58 / 8.31 5.73 / 14.28 13.30
GLM-5.3-Flash [61] Native† 12.19 / 32.41 8.89 / 27.03 4.95 / 18.28 59.71 / 91.13 27.51 / 61.05 7.13 / 16.33 8.07 / 19.79 18.35
Open-Vocabulary Segmentation Models
YOLOE [57] 6402 20.98 / 27.52 26.61 / 38.72 16.00 / 21.41 83.53 / 90.60 26.97 / 40.48 24.06 / 33.58 31.48 / 39.48 32.80
YOLO-World [9] 6402 23.97 / 40.03 36.94 / 60.44 29.62 / 42.69 61.46 / 80.40 27.67 / 45.20 24.21 / 37.42 29.93 / 49.26 33.40
Grounded SAM 2 [23] 10242 44.06 / 56.85 47.08 / 67.96 43.93 / 67.63 92.13 /98.6059.02 / 77.94 63.35 / 78.67 50.37 / 60.99 57.13
SAM 3 [7] 6442 51.59 / 68.19 55.27 / 80.53 36.98 / 60.11 89.06 / 95.75 60.11 / 81.55 60.92 / 82.67 61.31 / 75.75 59.32
SAM 3 [7] 10082 59.50 / 74.65 60.10 /81.0935.36 / 58.06 90.49 / 96.73 66.31 / 83.5774.07/89.0264.30 / 79.56 64.30
SAF (Ours)
SAF-YOLO26m-seg 6402 53.94 / 75.24 49.33 / 76.4349.45/ 74.45 92.25 / 97.28 70.29 / 88.75 57.48 / 80.86 62.28 / 79.39 62.14
SAF-DINOv3+M2F 6402 55.19 / 71.39 52.20 / 74.97 46.20 / 71.88 92.34 / 96.26 68.25 / 81.91 56.32 / 73.93 60.87 / 77.40 61.62
SAF-DINOv3+M2F ℓ≤102458.86 / 74.57 56.57 / 76.80 46.64 / 71.72 92.69 / 96.43 71.34 / 84.25 65.79 / 80.93 63.32 / 79.24 65.03
SAF-RF-DETR-XL 6242 64.01/79.51 60.14/ 80.46 48.39 /75.49 94.21/ 97.0675.33/90.0372.95 / 87.47 68.38/82.66 69.06
whereas SAF runs prompt-free. We report native input res-
olutions and provide box metrics in the supplement.
4.1. Generalization and Edge Deployment
Zero-shot accuracy across species and field conditions.
All four SAF configurations outperform Grounded SAM 2
and SAM 3 (644 2) in Table 1, reaching up to 69.06 aver-
age mask AP50:95 (+4.76 over SAM 3 at1008 2), with SAF-
RF-DETR-XL and SAF-DINOv3+M2F (ℓ≤1024) setting
the top marks. Across the seven benchmarks, SAF mod-
els rank first on six (five led by SAF-RF-DETR-XL, includ-
ing a +9.02 AP50:95 gain on StrawDI, and KFuji-Apple led
by SAF-YOLO26m-seg) while remaining competitive with
the 848M-parameter SAM 3 on Tomato. Beyond cross-
species generalization, SAF resists intra-class acquisition
shifts: moving from daylight MSU-Apple to nighttime
KFuji-Apple degrades SAM 3 (1008 2) by 24.74 AP 50:95,
whereas SAF-RF-DETR-XL cuts this drop by more than
half (to 11.75) and SAF-YOLO26m-seg maintains virtually
unchanged accuracy. Qualitatively (Figure 3), SAF-RF-
DETR-XL produces cleaner boundaries with fewer false
positives on unseen WebOOD fruits and harvested clusters.
Together, these results show that distilling intrinsic morpho-
logical priors yields far stronger perceptual invariance than
relying on generic text-prompted foundation models.
Mask precision separates SAF from VLMs.Evalu-
ated frontier VLMs trail even our lightest SAF model by at
least 14.32 average mask AP50:95 (Table 1). While GPT-6-
Astra reaches 84.74 AP 50:95 on relatively isolated Mango
fruits, its mask accuracy drops sharply to 37.84 on dense,
occluded Tomato canopies. Visually, VLMs suffer from se-
Table 2. TensorRT engine throughput on Jetson Orin NX 8GB.
Model Params (M) Input FP32 FPS FP16 FPS
SAM3 8481008 2 1.40 1.86
YOLO26n-seg 3.1640 2 121.75 253.75
YOLO26s-seg 11.4640 2 64.90 148.83
YOLO26m-seg 27.0640 2 28.79 69.22
DINOv3+M2F 116.1480×6403.33 5.51
RF-DETR-M 35.4432 2 23.04 66.87
RF-DETR-L 35.9504 2 16.21 49.51
RF-DETR-XL 38.1624 2 8.52 30.01
Figure 3. Qualitative comparison with SAF-RF-DETR-XL on (a)
unseen fruit types in WebOOD and (b) clustered blackcurrants.
vere false negatives and coarse contours in crowded clusters
(Figure 3a), confirming that broad semantic grounding can-
not replace dedicated instance boundary supervision.
5
0 1 5 10 Full
20
40
60
Mask AP50 : 95
13.98
17.83
64.05
59.04
68.36 (+4.31)
MSU (880 Images)
60.89
0 1 5 10 Full
20
40
60
9.97
15.80
67.23
48.49
69.81 (+2.58)
KFuji (774 Images)
60.79
0 1 5 10 Full
40
60
80
100
40.76
63.53
94.1794.21
96.29 (+2.12)
Mango (1,054 Images)
94.22
0 1 5 10 Full
20
40
60
80
24.03
31.27
81.36
76.33
82.83 (+1.47)
StrawDI (2,900 Images)
77.75
0 1 5 10 Full
0
20
40
60
80
2.63
3.70
69.3472.95
74.83 (+5.49)
Tomato (116 Images)
74.05
MSU KFuji MangoStrawDI TomatoWebTestWebOODOther-6
MSU
KFuji
Mango
StrawDI
Tomato
+5.0 +1.6 -7.3 -17.8 -23.3 -22.5 -13.9 -13.9
-9.1 +18.7 -11.7 -21.4 -51.0 -33.6 -27.7 -25.7
-9.0 -3.6 +0.0 -11.4 -46.4 -22.4 -16.1 -18.2
-30.4 -18.4 -9.4 +5.0 -32.2 -28.9 -22.9 -23.7
-12.7 -12.9 -12.2 -6.9 -3.6 -17.3 -9.0 -11.8
From Official Weights
MSU KFuji MangoStrawDI TomatoWebTestWebOODOther-6
MSU
KFuji
Mango
StrawDI
Tomato
+9.3 +8.3 +0.3 -3.8 -5.9 -2.2 -1.5 -0.8
+2.6 +21.3 -0.8 -6.0 -14.0 -5.9 -3.1 -4.6
+1.9 +8.2 +2.1 -2.7 -9.1 -0.8 +0.1 -0.4
-6.2 -9.4 -0.9 +6.5 -2.7 -4.9 -3.6 -4.6
-2.9 -3.1 -0.7 +0.9 +1.9 -0.2 -0.3 -1.0
From Our Weights
-60
-30
0
30
60
(a) Label Efficiency Across Five Target Domains
(b) Cross-Domain Performance After Full-Shot Adaptation
Boxed Cells : Adaptation Target.
Other-6
: Mean over Six Non-Target Tests. Positive = Improvement.
From Our Weights From Official Weights
Figure 4. Adaptation and retention with RF-DETR-XL. (a) Target AP 50:95 vs. labeled images (error bars: sample SD over five runs). (b)
Full-shot AP50:95 changes over the unadapted SAF checkpoint. Boxes mark adaptation targets; Other-6 averages non-target tests.
Pixel efficiency and real-time edge deployment.
Higher input resolution consistently improves dense fruit
segmentation, lifting DINOv3+M2F and SAM 3 by 3.41
and 4.98 average mask AP 50:95, respectively. However,
SAF models achieve strong accuracy without relying on
high-resolution inputs. Operating at640 2 (∼60% fewer pix-
els than SAM 3 at1008 2), SAF-YOLO26m-seg still out-
performs high-resolution SAM 3 on KFuji, Mango, and
StrawDI. On a Jetson Orin NX 8GB (Super Mode, Ta-
ble 2), heavy foundation backbones fall far short of real-
time speed, with SAM 3 and DINOv3+M2F reaching only
1.86 and 5.51 TensorRT [39] FP16 FPS. In contrast, all
SAF detectors run comfortably in real time: RF-DETR vari-
ants deliver 30.01 to 66.87 FPS (making SAF-RF-DETR-
XL over16×faster than SAM 3 with22×fewer parame-
ters), while YOLO26-seg scales up to 253.75 FPS. Addi-
tional edge benchmarks are detailed in the supplement.
4.2. Few-shot adaptation and retention
We evaluate SAF’s target adaptation and broad retention
across 1-, 5-, 10-, and full-shot settings (1-shot: one labeled
image) over five seeds, comparing RF-DETR-XL initialized
from SAF versus official pretrained weights (Figure 4; de-
tails in the supplement).
Label-efficient adaptation and higher performance
ceiling.SAF provides a strong domain-aligned ini-
tialization across all five agricultural benchmarks (Fig-
ure 4a). Without any target supervision, zero-shot SAF al-
ready matches the full-shot official-initialized specialist on
Mango and surpasses it by 3.61 AP 50:95 on dense Tomato
canopies. Under extreme label scarcity, fine-tuning on a sin-
gle image triggers rapid adaptation, enabling SAF to outper-
form 1-shot official baselines by +44.99 AP 50:95 on KFuji
and +43.06 on MSU. Crucially, these advantages persist as
target data scales: under full supervision, SAF initializa-
tion consistently raises the performance ceiling across all
five domains (+1.47 to +5.49 AP50:95 over official weights),
showing that agricultural pre-training improves both sample
efficiency and asymptotic accuracy.
Preserving cross-environment generalization after
specialization.Figure 4b compares post-adaptation
AP50:95 across all seven benchmarks relative to the same
unadapted SAF checkpoint. Standard full-shot train-
ing from official weights leaves severe cross-environment
deficits, trailing unadapted SAF by 11.8–25.7 AP 50:95 on
average across the six non-target benchmarks (Other-6). In
contrast, adapting from SAF improves every target domain
(+1.9 to +21.3 AP 50:95) while keeping the averageOther-
6reduction within just 0.4–4.6 AP 50:95. Notably, SAF
even enables positive cross-environment transfer: adapting
to MSU or Mango raises KFuji performance by +8.3 and
+8.2 AP50:95, respectively. Even under the sharpest cross-
species shift (KFuji to Tomato), SAF limits the off-target
drop to 14.0 AP 50:95 (versus 51.0 points below unadapted
SAF for the official-initialized specialist), demonstrating
strong resistance to catastrophic forgetting.
6
Table 3. Training-stage ablation. Seven-test average mask
AP50:95;✓/×: stage included/omitted. Column bests are bold.
Pre Mid Post YOLO26m DINOv3+M2F RF-XL
×✓×50.45 56.69 64.97
✓ ✓×50.93 62.72 65.01
× ×✓62.77 59.57 68.29
✓×✓62.32 63.33 68.83
×✓ ✓63.5459.96 68.89
✓ ✓ ✓62.1465.03 69.06
Figure 5. Data and model scaling. Seven-test average mask
AP50:95 for (a) WebFruits Mid, (b) PaperFruits Post with three
initializations, and (c) model configurations.
4.3. Supervision budgets and model scale
Table 3 compares full-data training routes across architec-
tures, while Figure 5 examines how performance scales
with supervision budgets and model capacity.
Dense cultivation supervision anchors final accuracy.
Across all architectures, incorporating PaperFruits (Post)
yields the largest single-stage improvement, boosting re-
sults by 11.21, 2.31, and 4.05 average mask AP 50:95 on
YOLO26m, DINOv3+M2F, and RF-DETR-XL, respec-
tively (Table 3). While WebFruits (Mid) provides broad
appearance coverage, PaperFruits supplies dense orchard
scenes and refined instance boundaries essential for practi-
cal fruit segmentation. Upstream stages (PreandMid) con-
tribute complementary priors whose impact varies by archi-
tecture: DINOv3+M2F gains 5.46 average mask AP 50:95
over Post-only training with all three stages (reaching
65.03 average mask AP 50:95), and RF-DETR-XL peaks at
69.06 average mask AP50:95 with the full pipeline, whereas
YOLO26m achieves its strongest result (63.54 average
Table 4. Synthetic augmentation with 320 human-labeled images.
+Syn adds 31,308 automatically labeled frames to the same origi-
nals. All entries are mask AP 50:95. Average: equal-weight seven-
test average;∆: +Syn−Human.
Initialization average mask AP50:95 ∆AP 50:95
Human +Syn Average KFuji WebOOD
Official 58.9063.75+4.85 +4.26 +8.41
Pre 62.6263.65+1.04 +4.62 +0.32
Pre + Mid66.9164.69−2.22 +2.47 +0.68
mask AP50:95) with Mid + Post.
Upstream supervision dramatically reduces down-
stream annotation needs.While upstream stages yield
modest gains when Post data are abundant, they act as pow-
erful annotation multipliers in label-scarce regimes (Fig-
ure 5a–b). During WebFruits scaling (Figure 5a), SA-
1B Pre provides a +2.12 average mask AP 50:95 head start
at 0.1% data, after which both Pre and official-initialized
curves steadily climb and converge at 100% WebFruits.
When transitioning to PaperFruits with a 0.1% label pool
(Figure 5b), adding full Mid on top of Pre raises perfor-
mance from 55.28 to 66.99 average mask AP 50:95 (+11.71
across seven benchmarks), far surpassing the 16.17 from
official weights. Pairing full Mid with just 1% PaperFruits
reaches 68.20 average mask AP 50:95, recovering 98.8% of
the full-data reference (69.06) with two orders of magnitude
fewer manual labels, as Mid absorbs most domain variance
and leaves Post to calibrate fine boundaries.
Model capacity governs the ceiling of multi-stage su-
pervision.Figure 5c shows accuracy improving with model
scale within each family, compounded by higher native in-
put resolutions in larger RF-DETR variants. With abun-
dant Post supervision, compact models (SAF-YOLO26n/s
and SAF-RF-DETR-M) saturate quickly and gain modestly
from upstream training, whereas higher-capacity models
(SAF-DINOv3+M2F and SAF-RF-DETR-XL) possess the
representational bandwidth to retain web-scale priors and
benefit from the full three-stage recipe. Consequently, com-
pact students suit edge deployment when curated Post data
are available, while larger models unlock the full value of
large-scale upstream supervision.
4.4. Synthetic Augmentation
To test whether generative synthesis alleviates label
scarcity, we build a prompt-guided Seedance 2.5 video
pipeline (Figure 6a; details in supplement). From each
seed image, we synthesize 20 five-second clips (100 frames
at 1 fps) varying lighting, canopy density, viewpoint, and
ripeness. Using 320 seeds from 16 PaperFruits subsets (20
each) yields 32,000 frames; retaining fruit-present frames
adds 31,308 SAM 3-labeled frames to 320 human-labeled
images, keeping all five test domains held out.
7
Figure 6. Video augmentation and held-out KFuji predictions. (a)
Generated lighting, ripeness, and viewpoint variations. (b) Predic-
tions trained on 320 human-labeled images with or without syn-
thetic frames (conf. 0.30, IoU 0.50).
As shown in Table 4, synthetic expansion delivers broad
gains when real fruit supervision is limited, lifting aver-
age mask AP50:95 by +4.85 from official weights (58.90 to
63.75) and by +1.04 under SA-1BPreinitialization (62.62
to 63.65), while suppressing canopy false positives on held-
out KFuji (Figure 6b; FP drops from 12 to 2 as TP rises
from 2 to 5). Notably, KFuji and WebOOD benefit con-
sistently across all three initializations, still gaining +2.47
and +0.68 AP 50:95 even afterPre+Mid(where overall av-
erage mask AP 50:95 dips by 2.22 once 570K+ real images
have covered common viewpoint and ripeness variations).
Unlike agricultural collections captured under routine illu-
mination, nighttime KFuji and photography-rich WebOOD
exhibit strong backlighting, warm golden/dusk casts, and
deep-shadow contrast, precisely the lighting regimes tar-
geted by our video synthesis. Generative augmentation thus
acts as a versatile booster in scarce-data regimes and a tar-
geted complement for underrepresented lighting conditions.
4.5. Cross-Domain Recipe Validation
To verify that ourPre–Mid–Postparadigm generalizes be-
yond agriculture, we apply the same recipe, with zero fruit
data or fruit-trained weights, to underwater salient instances
(USIS10K [33]), microscopy cells (LIVECell [13]), and oc-
cluded humans (CIHP [19] training, OCHuman [62] test-
ing) using RF-DETR-XL (37.79M parameters). In each
task,Midreplaces manual labeling with automated pseudo-
masks (Qwen-3.8-max [46]+SAM 2.1 [47] for USIS10K
and SAM 3 [7] for the other tasks), whilePostuses only
1% human labels on USIS10K/LIVECell (74/32 images)
and 0.1% on humans (28 images); selection-label budgets
Table 5. Cross-domain validation of the SAF training recipe.
Limited-label Post uses 1% GT for USIS10K/LIVECell and 0.1%
for humans. All metrics are mask AP, following the authors’ orig-
inal evaluation settings.†Author-reported results [13, 29, 33].
Model / route USIS10K LIVECell OCHuman
AP50:95 AP50 AP50:95 AP50 AP50:95 AP50
Full supervision
Authors’ model† 64.30 84.9047.89— 34.70 —
Post68.74 87.5147.79 80.4551.14 64.95
Limited human labels
Post 52.31 68.89 16.61 33.29 37.73 50.26
Pre + Post 57.41 74.95 33.09 60.62 35.11 53.50
Pre + Mid 61.25 81.24 42.28 74.26 35.58 47.15
Mid + Post 62.31 81.74 45.02 78.4648.23 62.79
Pre + Mid + Post64.18 83.57 45.79 79.2846.86 61.94
and protocols are detailed in the supplement.
As shown in Table 5, automatedMidsupervision is con-
sistently the primary driver of label efficiency. Even where
SA-1BPreyields mixed returns (helping USIS10K and
LIVECell, yet slightly trailing on OCHuman where public
COCO weights already encode human priors), addingMid
alone lifts limited-labelPostby +10.00, +28.41, and +10.50
AP50:95 on USIS10K, LIVECell, and OCHuman, respec-
tively, and reaches 61.25 and 42.28 AP 50:95 on USIS10K
and LIVECell with 0% human labels (Pre+Mid). Paired
with minimalPostcalibration, our compact 37.79M model
recovers 93.4%–95.8% of its 100%-GT upper bound across
all benchmarks (64.18 vs. 68.74, 45.79 vs. 47.79, and 48.23
vs. 51.14 AP 50:95), while matching or surpassing larger
100%-supervised author baselines on USIS10K (64.30,
∼701M) and OCHuman (34.70, 80.85M).
5. Conclusion
In this work, we demonstrated that achieving foundation-
model-level generalization in specialized visual domains
does not require deploying massive, prompt-dependent
models at runtime. Through progressivePre-Mid-Postsu-
pervision across WebFruits and PaperFruits,SegAnyFruit
(SAF) distills broad foundation priors into compact,
prompt-free segmenters capable of real-time onboard in-
ference. Our findings highlight three practical takeaways
for domain-specific perception: (1) learning a unified visual
concept across diverse environments yields strong zero-shot
transfer while serving as a resilient initialization that curbs
catastrophic forgetting during local adaptation; (2) large-
scale upstream supervision acts as a potent annotation mul-
tiplier, recovering near-full-data accuracy with as little as
1% human-verified domain labels; and (3) this progressive
distillation recipe extends naturally beyond agriculture to
other data-scarce segmentation tasks such as underwater
8
and cellular imagery.
Limitations and future work.Our current framework
has three primary limitations. First, automatically generated
pseudo-masks still retain errors in boundary quality and in-
stance granularity. Second, although the benefits of generic
mask pretraining vary across architectures and downstream
tasks, we have not yet established when it is most effec-
tive. Third, SAF treats fruit as a unified visual concept
and does not distinguish individual species. In future work,
we will focus on improving automated annotation quality
for sharper boundaries and consistent granularity, extend-
ing SAF to fine-grained species recognition, and conduct-
ing larger-scale studies across domains and architectures to
clarify how training stages should be best combined.
References
[1] Manya Afonso, Hubert Fonteijn, Felipe Schadeck Fiorentin,
Dick Lensink, Marcel Mooij, Nanne Faber, Gerrit Polder,
and Ron Wehrens. Tomato fruit detection and counting in
greenhouses using deep learning.Frontiers in Plant Science,
11:571299, 2020. 4
[2] Shuai Bai et al. Qwen3-VL technical report, 2025. 2
[3] Suchet Bargoti and James Underwood. Deep fruit detection
in orchards. InIEEE International Conference on Robotics
and Automation, 2017. 2, 3
[4] Caner Beldek, Zoe Cobbler, Emre Sariyildiz, Gursel Alici,
and Son Lam Phung. Berremangra orange dataset. Zenodo,
version 1, 2026. 4
[5] Md. Hasan Imam Bijoy, Syeda Zarin Tasnim, Syed Ali
Awsaf, and Md. Zahid Hasan. FruitVision: A benchmark
dataset for fresh, rotten, and formalin-mixed fruit detection.
Data in Brief, 61:111752, 2025. 4
[6] Alessandro Carella, Baptiste Paul Ernest Lucas, Safouane
El Ghazouali, Pedro Tomas Bulacio Fischer, Roberto
Massenti, Francesca Venturini, Umberto Michelucci, and
Riccardo Lo Bianco. Large-scale orange fruit dataset for lo-
calization, classification and ripening assessment under vary-
ing environments.Computers and Electronics in Agriculture,
248:111833, 2026. 2
[7] Nicolas Carion, Laura Gustafson, Yuan-Ting Hu, Shoub-
hik Debnath, Ronghang Hu, Didac Suris, Chaitanya Ryali,
Kalyan Vasudev Alwala, Haitham Khedr, Andrew Huang,
Jie Lei, Tengyu Ma, Baishan Guo, Arpit Kalla, Markus
Marks, Joseph Greer, Meng Wang, Peize Sun, Roman
R¨adle, Triantafyllos Afouras, Effrosyni Mavroudi, Kather-
ine Xu, Tsung-Han Wu, Yu Zhou, Liliane Momeni, Rishi
Hazra, Shuangrui Ding, Sagar Vaze, Francois Porcher, Feng
Li, Siyuan Li, Aishwarya Kamath, Ho Kei Cheng, Piotr
Doll´ar, Nikhila Ravi, Kate Saenko, Pengchuan Zhang, and
Christoph Feichtenhofer. SAM 3: Segment anything with
concepts, 2026. 2, 3, 5, 8
[8] Bowen Cheng et al. Masked-attention mask transformer for
universal image segmentation. InIEEE/CVF Conference on
Computer Vision and Pattern Recognition, 2022. 4
[9] Tianheng Cheng et al. YOLO-World: Real-time open-
vocabulary object detection. InIEEE/CVF Conference on
Computer Vision and Pattern Recognition, 2024. 2, 5
[10] Mehdi Cherti, Romain Beaumont, Ross Wightman, Mitchell
Wortsman, Gabriel Ilharco, Cade Gordon, Christoph Schuh-
mann, Ludwig Schmidt, and Jenia Jitsev. Reproducible
scaling laws for contrastive language-image learning. In
IEEE/CVF Conference on Computer Vision and Pattern
Recognition, pages 2818–2829, 2023. 2
[11] Pengyu Chu, Zhaojian Li, Kaixiang Zhang, Dong Chen,
Kyle Lammers, and Renfu Lu. O2RNet: Occluder-occludee
relational network for robust apple detection in clustered
orchard environments.Smart Agricultural Technology, 5:
100284, 2023. 4
[12] DeepSeek-AI. DeepSeek-V4.1-Flash: Pushing the limits of
KV cache compression, 2026. 5
[13] Christoffer Edlund, Timothy R. Jackson, Nabeel Khalid,
Nicola Bevan, Timothy Dale, Andreas Dengel, Sheraz
Ahmed, Johan Trygg, and Rickard Sj ¨ogren. LIVECell—a
large-scale dataset for label-free live cell segmentation.Na-
ture Methods, 18(9):1038–1045, 2021. 8
[14] Chengxiang Fan, Muzhi Zhu, Hao Chen, Yang Liu, Weijia
Wu, Huaqi Zhang, and Chunhua Shen. DiverGen: Improv-
ing instance segmentation by learning wider data distribution
with more diverse generative data. InIEEE/CVF Conference
on Computer Vision and Pattern Recognition, pages 3986–
3995, 2024. 3
[15] Zhenghao Fei, Alex Olenskyj, Brian N. Bailey, and Mason
Earles. Enlisting 3D crop models and GANs for more data
efficient and generalizable fruit detection.arXiv preprint
arXiv:2108.13344, 2021. 1, 2, 3
[16] Flickr. Flickr API.https://www.flickr.com/
services/api/, n.d. Accessed November 16, 2026. 3
[17] GBIF Secretariat. GBIF occurrence API.https :
/ / techdocs . gbif . org / en / openapi / v1 /
occurrence, n.d. Accessed November 16, 2026. 3
[18] Jordi Gen ´e-Mola, Ver´onica Vilaplana, Joan R. Rosell-Polo,
Josep-Ramon Morros, Javier Ruiz-Hidalgo, and Eduard Gre-
gorio. KFuji RGB-DS database: Fuji apple multi-modal im-
ages for fruit detection with color, depth and range-corrected
IR data.Data in Brief, 25:104289, 2019. 4
[19] Ke Gong, Xiaodan Liang, Yicheng Li, Yimin Chen, Ming
Yang, and Liang Lin. Instance-level human parsing via part
grouping network. InEuropean Conference on Computer
Vision, pages 770–785, 2018. 8
[20] Nicolai H ¨ani, Pravakar Roy, and V olkan Isler. MinneApple:
A benchmark dataset for apple detection and segmentation.
IEEE Robotics and Automation Letters, 5(2):852–858, 2020.
3
[21] Saihui Hou, Yushan Feng, and Zilei Wang. VegFru: A
domain-specific dataset for fine-grained visual categoriza-
tion. InIEEE International Conference on Computer Vision,
pages 541–549, 2017. 3, 4
[22] Chenxiao Huang, Linran He, Wentao Huang, and Xiaoshuan
Zhang. TMAF-YOLO: A lightweight model for in situ detec-
tion of tomato maturity and defective fruits in greenhouses.
Sensors, 26(15):4950, 2026. 2
[23] IDEA Research. Grounded SAM 2: Ground and track
anything in videos.https://github.com/IDEA-
9
Research/Grounded- SAM- 2, n.d. Official project
repository. Accessed November 17, 2026. 5
[24] iNaturalist. iNaturalist API.https : / / api .
inaturalist.org/v1/docs/, n.d. Accessed Novem-
ber 16, 2026. 3
[25] Ariful Islam, Sadia Afroz Oishi, Md. Maksudul Hasan,
Sajeeb Kumar Ray, Md. Anwar Hossain, Naima Islam, and
Anomik Kumar. MangoFruitBD: A bounding-box anno-
tated image dataset for detecting healthy and diseased mango
fruits in bangladeshi orchards. Mendeley Data, V1, 2026. 4
[26] Weikuan Jia, Jie Liu, Yuqi Lu, Qiaolian Liu, Ting Zhang, and
Xishang Dong. Polar-Net: Green fruit instance segmentation
in complex orchard environment.Frontiers in Plant Science,
13:1054007, 2022. 1, 2
[27] Glenn Jocher, Jing Qiu, Mengyu Liu, Shuai Lyu, Fatih Ca-
gatay Akyon, and Muhammet Esat Kalfaoglu. Ultralytics
YOLO26: Unified real-time end-to-end vision models, 2026.
2, 4
[28] Alexander Kirillov et al. Segment anything. InIEEE/CVF
International Conference on Computer Vision, 2023. 2, 4
[29] Constantin Kolomiiets, Miroslav Purkrabek, and Jiri Matas.
SAM-pose2seg: Pose-guided human instance segmentation
in crowds. In29th Computer Vision Winter Workshop, 2026.
8
[30] Tingting Leng, Yuxin Zhang, Yamin Fan, Huajuan Gao,
Binhong Zhou, and Jiong Mu. A lightweight shape-aware
YOLO network for field strawberry maturity detection under
complex orchard conditions.Agronomy, 16(11):1074, 2026.
2
[31] Jiajia Li, Kyle Lammers, Xunyuan Yin, Xiang Yin, Long
He, Jun Sheng, Renfu Lu, and Zhaojian Li. MetaFruit meets
foundation models: Leveraging a comprehensive multi-fruit
dataset for advancing agricultural foundation models.Com-
puters and Electronics in Agriculture, 231:109908, 2025. 2
[32] Weijia Li. Litchi-UA V: A UA V-based litchi fruit detection
dataset for precision agriculture. Zenodo, 2026. 4
[33] Shijie Lian, Ziyi Zhang, Hua Li, Wenjie Li, Laurence Tian-
ruo Yang, Sam Kwong, and Runmin Cong. Diving into
underwater: Segment anything model guided underwater
salient instance segmentation and a large-scale dataset. In
Proceedings of the 41st International Conference on Ma-
chine Learning, pages 29545–29559, 2024. 8
[34] Tsung-Yi Lin et al. Microsoft COCO: Common objects in
context. InEuropean Conference on Computer Vision, 2014.
4
[35] Shilong Liu et al. Grounding DINO: Marrying DINO with
grounded pre-training for open-set object detection. InEu-
ropean Conference on Computer Vision, 2024. 2
[36] Baptiste Lucas, Alessandro Carella, Safouane El Ghazouali,
Pedro Bulacio, Roberto Massenti, Francesca Venturini, Um-
berto Michelucci, and Riccardo Lo Bianco. Oranges in the
field. Mendeley Data, version 1, 2025. 4
[37] Sandro Augusto Magalh ˜aes, Germano Moreira, Filipe Neves
dos Santos, and M ´ario Cunha. AgRobTomato dataset:
Greenhouse tomatoes with different ripeness stages. Zen-
odo, version 1, 2021. 4
[38] Matthias Minderer, Alexey Gritsenko, and Neil Houlsby.
Scaling open-vocabulary object detection. InAdvances in
Neural Information Processing Systems, 2023. 2
[39] NVIDIA. NVIDIA TensorRT documentation.https:
//docs.nvidia.com/deeplearning/tensorrt/
10.x.x/, 2026. Accessed November 17, 2026. 6
[40] OpenAI. GPT-6 Astra model.https://developers.
openai.com/api/docs/models/gpt- 6- astra,
2026. Accessed November 17, 2026. 5
[41] OpenAI. Introducing GPT-6 Sol and Luna.https://
openai.com/index/introducing-gpt-6-sol-
and-luna/, 2026. Accessed November 17, 2026. 5
[42] OpenAI. Introducing GPT-6.1 Sol.https://openai.
com/index/introducing- gpt- 6- 1- sol/, 2026.
Accessed November 17, 2026. 5
[43] OpenAI. Introducing GPT-6.1 Terra.https://openai.
com / index / introducing - gpt - 6 - 1 - terra/,
2026. Accessed November 17, 2026. 5
[44] Manthan Patel, Jonas Frey, Mayank Mittal, Fan Yang,
Alexander Hansson, Amir Bar, Cesar Cadena, and Marco
Hutter. DeFM: Learning foundation representations from
depth for robotics, 2026. Version 3. Accepted for publica-
tion in IEEE Transactions on Robotics. 3
[45] Isaac P ´erez-Borrero, Diego Mar ´ın-Santos, Manuel E.
Geg´undez-Arias, and Estefan ´ıa Cort ´es-Ancos. A fast and
accurate deep learning method for strawberry instance seg-
mentation.Computers and Electronics in Agriculture, 178:
105736, 2020. 4
[46] Qwen Team. Qwen3.8-Max.https : / / qwen . ai /
blog?id=qwen3.8, 2026. Accessed November 17, 2026.
8
[47] Nikhila Ravi, Valentin Gabeur, Yuan-Ting Hu, Ronghang
Hu, Chaitanya Ryali, Tengyu Ma, Haitham Khedr, Roman
R¨adle, Chloe Rolland, Laura Gustafson, Eric Mintun, Junt-
ing Pan, Kalyan Vasudev Alwala, Nicolas Carion, Chao-
Yuan Wu, Ross Girshick, Piotr Doll´ar, and Christoph Feicht-
enhofer. SAM 2: Segment anything in images and videos. In
The Thirteenth International Conference on Learning Rep-
resentations, 2025. 2, 8
[48] Shaoqing Ren, Kaiming He, Ross Girshick, and Jian Sun.
Faster R-CNN: Towards real-time object detection with re-
gion proposal networks. InAdvances in Neural Information
Processing Systems, 2015. 2
[49] Tianhe Ren et al. DINO-X: A unified vision model for open-
world object detection and understanding, 2024. 2
[50] Isaac Robinson, Peter Robicheaux, Matvei Popov, Deva Ra-
manan, and Neehar Peri. RF-DETR: Neural architecture
search for real-time detection transformers. InInternational
Conference on Learning Representations, 2026. 4
[51] Inkyu Sa et al. Deepfruits: A fruit detection system using
deep neural networks.Sensors, 16(8), 2016. 2
[52] Thiago T. Santos, Leonardo L. de Souza, Andreza A. dos
Santos, and Sandra Avila. Grape detection, segmentation and
tracking using deep neural networks and three-dimensional
association.Computers and Electronics in Agriculture, 170:
105247, 2020. 3, 4
[53] Schram03. fruits. Hugging Face dataset repository, n.d. Ac-
cessed November 16, 2026. 4
10
[54] Oriane Sim ´eoni et al. DINOv3, 2025. 3, 4
[55] Alessandra Tafuro, Adeayo Adewumi, Soran Parsa, Ghala-
mzan E. Amir, and Bappaditya Debnath. Strawberry pick-
ing point localization ripeness and weight estimation. In
IEEE International Conference on Robotics and Automation,
pages 2295–2302, 2022. 4
[56] TJIET. FruitBench: A multimodal benchmark for fruit
growth understanding. Hugging Face dataset repository, n.d.
Accessed November 16, 2026. 4
[57] Ao Wang et al. YOLOE: Real-time seeing anything. In
IEEE/CVF International Conference on Computer Vision,
2025. 2, 5
[58] Yanan Wang, Zhenghao Fei, Ruichen Li, and Yibin Ying.
Learn from foundation model: Fruit detection model without
manual annotation.Pattern Recognition, 174:112799, 2026.
2, 3
[59] Qizhe Xie, Minh-Thang Luong, Eduard Hovy, and Quoc V .
Le. Self-training with noisy student improves ImageNet clas-
sification. InIEEE/CVF Conference on Computer Vision and
Pattern Recognition, 2020. 3
[60] Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi
Feng, and Hengshuang Zhao. Depth Anything: Unleashing
the power of large-scale unlabeled data. InIEEE/CVF Con-
ference on Computer Vision and Pattern Recognition, pages
10371–10381, 2024. 3
[61] Z.ai. GLM-5.3-Flash: Official model card.https :
//huggingface.co/zai- org/GLM- 5.3- Flash,
2026. Accessed November 17, 2026. 5
[62] Song-Hai Zhang, Ruilong Li, Xin Dong, Paul L. Rosin, Zixi
Cai, Xi Han, Dingcheng Yang, Hao-Zhi Huang, and Shi-Min
Hu. Pose2Seg: Detection free human instance segmentation.
InIEEE/CVF Conference on Computer Vision and Pattern
Recognition, 2019. 8
[63] Xinyu Zhang, Sawut Mamat, Xiaohuang Liu, Jiufen Liu,
Run Liu, Guangjie Wu, Ping Zhu, Hongyu Li, Min Ma,
and Xiaotong Liu. MSRRT-DETR: A high-precision ap-
ple detection method with strong cross-domain generaliza-
tion capability in complex orchard scenes.PLOS ONE, 21
(3):e0342854, 2026. 2
[64] Zhuoyang Zhang, Han Cai, and Song Han. EfficientViT-
SAM: Accelerated segment anything model without perfor-
mance loss. InProceedings of the IEEE/CVF Conference on
Computer Vision and Pattern Recognition Workshops, pages
7859–7863, 2024. 2
Appendix
Additional supplementary material will be provided in a
forthcoming revision.
11
论文原文与出版信息
arXiv