- 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
-
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.
-
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
-
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
-
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?