Answers are grounded in this paper and cite their sources.
Environmental variation changes the background appearance against which weeds must be detected. We propose metadata-guided appearance-shift consistency (MASC), a training framework that uses environmental metadata and soil-colour statistics to vary background appearance. Original and transformed images share ground-truth detection supervision, with an additional loss encouraging stable predictions across the two views. The YOLO11n inference network remains unchanged. Evaluation used 8034 CropAndWeed images with acquisition-group-disjoint training, development and test partitions, followed by direct transfer to 3SeasonWeedDet10 without target-domain training or tuning. At a common confidence threshold of 0.25, external weed recall increased from 45.05 ± 4.04% to 49.97 ± 4.51%. Weed average precision at intersection over union 0.5 (AP50) increased from 41.62 ± 2.60% to 44.21 ± 4.52% (mean ± sample standard deviation). Recall, F1 and AP50 improved in all four seed pairs on the complete external release. The mean recall gain remained 4.31 percentage points after excluding all exact-duplicate groups. On the principal source partition, recall was 59.17 ± 1.40% versus 58.81 ± 2.47%, whereas mean all-class mAP50 increased from 66.05% to 67.88%. An alternative partition favoured MASC by 3.92 recall points for one seed. Development-selected oversampling with shifted-view supervision increased source recall by 2.62 points in a single-seed comparison. MASC improved weed detection on the evaluated external dataset while retaining baseline inference complexity, supporting its effectiveness for cross-dataset transfer.
No takes yet. Share an insight, caveat, or question.
Li et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: