acceptodds
Withdrawn from ICLR 2027

AgroGround: Multi-Granularity Grounded Recognition in Agriculture

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Agricultural visual models are generally evaluated for either recognition or localization, but reliable diagnosis requires identifying what is present and localizing the visual evidence that supports the prediction. Agricultural visual question answering (VQA) datasets contain rich semantic labels but often lack annotations linking them to image regions. Adding these annotations by hand is costly at scale. We introduce AgroGround, a large-scale dataset for grounded agricultural recognition, where models identify plant diseases and other agricultural targets and localize their corresponding image regions. We develop an automated annotation pipeline that uses semantic labels from eight existing agricultural VQA datasets to annotate both fine-grained regions such as disease lesions and whole objects, producing 794,850 instruction examples. Healthy images provide negative supervision for disease-related queries, teaching the model to return empty predictions. We train a shared vision-language model through supervised fine-tuning, combining known-target grounding instructions with instructions that require both target recognition and localization. We evaluate predicted identities, regions, their joint correctness, and healthy-image abstention on 1,480 human-verified images disjoint from all training data by file path and perceptual hash. Grounding-only fine-tuning reduces recognition accuracy from 51.8% to 29.1%, and adding recognition-and-localization instructions raises it to 72.6%. With positive images and spatial annotations held fixed, combining the two instruction formats raises joint accuracy from 19.2% to 43.3% with comparable grounding performance. Healthy negatives raise abstention on healthy images to 95.0%, and reinforcement learning improves lesion-level grounding. The resulting 2B model achieves higher grounding F1 than its annotation teacher on both our benchmark and the external PlantSeg test set. AgroGround establishes a benchmark for grounded agricultural recognition, measuring joint correctness of target identity and localization, as well as abstention on healthy images.

Primary area:
datasets and benchmarks
Keywords:
visual grounding, weak supervision, plant disease, agricultural vision-language models, grounded disease recognition, pseudo-labels, dataset, GRPO, existence-aware grounding, Qwen3-VL

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Fm97OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

AgroGround introduces a weakly supervised training corpus and a human-verified benchmark for agricultural grounded recognition. Its pipeline converts labels from eight agricultural VQA datasets into 794,850 lesion- and object-level instruction examples using GroundingDINO, SAM2, and geometric filtering. Qwen3-VL models are fine-tuned with mixed known-target grounding and four-candidate recognition-and-localization instructions, healthy negatives, and optional GRPO refinement using the GSPO loss and pseudo-label rewards. On a 1,480-image overlap-audited benchmark, mixed instructions improve closed-set recognition and joint correctness, while healthy negatives enable healthy-image abstention. RL improves lesion localization at a cost to recognition and disease-given healthy-image abstention. PlantSeg provides external evidence of known-target localization transfer. The main contribution is the data resource and controlled analysis of these capability trade-offs.

Critical strengths
  1. With images, spatial annotations, and the training recipe fixed, mixed instructions increase recognition accuracy from 29.1% to 72.6% and joint accuracy from 19.2% to 43.3%, while preserving comparable known-target grounding. The candidate-position analysis also identifies a prompt-copying failure mode, illustrating why accurate localization alone does not establish correct target recognition.

  2. The corpus adds lesion- and object-level supervision to existing agricultural VQA resources. Human-verified references, overlap audits, and external localization testing on PlantSeg strengthen the empirical contribution. Transparent reporting of annotation noise, hard-negative failures, and capability trade-offs makes the study informative beyond a pipeline demonstration.

Critical weaknesses
  1. The benchmark uses a single annotator with detector proposals visible. Although most final boxes were newly drawn, 36.8% retain proposals largely unchanged, and no independent agreement study or blinded audit is reported (Apps. C and R).Moreover, the negative share and headline RL checkpoint were chosen after benchmark inspection, while the main supervised 2B result uses the best of three seeds (Sec. 3.3; Table 8). The disclosed alternatives are helpful, but these choices weaken the separation between model selection and final evaluation, particularly for smaller performance differences. 30368_AgroGround_Multi_Granula

  2. On diseased images queried with an absent same-crop disease, abstention is only 15.2% for the supervised model and 7.2% after RL, yet wrong-target queries remain a secondary diagnostic (App. F).External evaluation is limited to known-target localization on PlantSeg, leaving cross-dataset recognition, joint correctness, and target-specific abstention untested (App. O).Since the main benchmark draws from the same source datasets as training, image-level deduplication does not establish source-held-out generalization. My concern is coverage and external validation of the complete task, not simply the number of external datasets.

Questions for authors

Q1 — Which selection rules were fixed before benchmark evaluation? Please clarify the respective roles of benchmark results and pseudo-label validation rewards in checkpoint selection (Sec. 3.3; App. J). Could the existing three-seed mean and standard deviation replace the best-seed supervised headline result? Q2 — Is an independently annotated or proposal-blind subset available to assess agreement in instance coverage, granularity, and box placement? Does the student–teacher ranking remain stable under such independent references? Q3 — Could the existing wrong-target evaluation become a primary condition with paired present- and absent-target queries? Is there external or source-held-out evidence for recognition, joint correctness, and target-specific abstention beyond known-target localization on PlantSeg?

Reviewer_1yyUOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes AgroGround for agricultural grounded recognition, combining automatic spatial annotation with shared VLM training. GroundingDINO, SAM2, and geometry filtering convert labels from eight agricultural VQA datasets into 794,850 lesion- and object-level instructions. Qwen3-VL training mixes known-target grounding, recognition among four candidates, and healthy-image negatives. Optional GRPO with GSPO loss uses rewards computed against the automatic boxes. Extensive experiments on the AgroGround benchmark and PlantSeg demonstrate the framework's effectiveness in agricultural target localization and its ability to transfer across datasets.

Critical strengths
  1. This paper studies category recognition together with the spatial evidence supporting each prediction. Introducing healthy images also makes disease absence part of the learning problem, providing a broader basis for studying reliable agricultural recognition.

2.The experiments examine recognition and localization under different training settings, providing evidence about how supervision affects both capabilities. Evaluation on the independently annotated PlantSeg dataset further demonstrates the method's effectiveness in agricultural localization and its ability to transfer across datasets.

Critical weaknesses
  1. Insufficient novelty. Using labeled candidate-selection examples is a standard form of supervised task mixing. Adding weighted task rewards to the existing GRPO/GSPO procedure does not by itself introduce a new learning approach. The proposed changes therefore amount to small modifications of established training methods, with insufficient algorithmic novelty.

2.Insufficient evidence for joint RL gains. The reported RL configurations do not demonstrate a joint benefit for recognition and localization. Joint accuracy is 44.8% after SFT and 43.1% after RL (Table 2), and recognition falls from 74.7% to 69.2%. Increasing the recognition reward yields 44.1% joint accuracy (Table 3), still below SFT. The main RL model is reported after about 1.2 epochs, while the ablations use one epoch, and there is no comparison with continued SFT at the same image resolution and sampling strategy (Appendices D and J). The current experiments therefore do not convincingly show that the proposed RL design is responsible for the localization gains or offers an advantage over further supervised training.

Questions for authors
  1. Please clarify the novelty of the proposed training strategy and its fundamental distinctions in design rationale from standard multi-task SFT followed by GRPO/GSPO with weighted rewards.

  2. Does the claim that grounding-only training erases recognition concern visual discrimination or the way the model answers candidate prompts? Which reported result separates these interpretations, given the mixed model's explicit candidate-selection supervision?

  3. What input resolutions are actually used after preprocessing in SFT and RL, and how often are dense-lesion images used in each? Could the authors compare RL with continued SFT using the same amount of training, image resolution, and sampling strategy?

  4. Why does the recognition-weighted RL model have higher grounding F1 but lower joint accuracy than SFT (44.1% versus 44.8% in Table 3)? Do the existing predictions show whether the localization gains occur on images whose disease is also correctly recognized?

Reviewer_bwq2OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes AgroGround: a collection of 8 agricultural VQA datasets that are extended by automatically annotating the samples with bounding boxes and segmentation masks. The result is an agricultural VQA dataset with 794,850 instruction samples. 1480 samples are human verified for benchmarking while the remainder of the dataset is designated for training. Negative samples (i.e. an image of a plant without any disease) are highlighted as an important characteristic. The paper fine-tunes and evaluates a VLM on AgroGround.

Critical strengths

Human verification. The paper describes that human verification is done on 1480 samples. This is a nice contribution that provides a useful dataset for the computer vision/agricultural community.

Critical weaknesses

Presentation. The paper is incomprehensible. It frequently provides dense explanations that are unnecessary to understand the core contribution. Numbers are presented prior to the necessary context to understand them. Sentences are often unconnected. There are generally all the telltale signs of extremely heavy LLM-usage. That wouldn’t necessarily be a problem if the content was comprehensible, but it is not.

Low significance. The main contribution is AgroGround. It is collected from existing agricultural VQA datasets and it is automatically annotated. This does not provide much value for other researchers or practitioners. Neither does it provide new insights. Likewise, the fine-tuning on AgroGround does not yield algorithmic novelties nor new insights.

Questions for authors

What is the value and new knowledge other researchers or practitioners are expected to draw from this work?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.