Research map / Detection and segmentation / Open-set and few-shot detection
Towards Efficient and Intelligent Laser Weeding: Method and Dataset for Weed Stem Detection
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces the Weed Stem Detection (WSD) dataset (7,161 field images, 11,151 weed instances with stem points) and an end-to-end YOLOv7-based detector with an added stem-coordinate regression head plus semi-supervised teacher-student learning for laser weeding.
Problem
Laser weeding needs the laser aimed at the weed stem root, but standard weed detectors localize bounding boxes (judged by mAP) whose centers are often far from the actual stem, causing wasted shots and risk to crops.
Approach
Add a stem coordinate regression branch to an object detector so crops, weeds and weed stem locations are predicted jointly end-to-end, evaluated with Euclidean distance to stem rather than mAP; extend with semi-supervised learning using pseudo-labels filtered by confidence and cosine similarity to ground-truth weed embeddings.
Key results
- Stem regression with detection reduces stem Euclidean distance from 4.3306 (regression only) to 2.4838, and to 2.1485 with semi-supervised learning.
- Compared with the vanilla YOLOv7 detection method, the proposed method improves weeding accuracy from 75.37% to 80.42% (+6.7% relative stated) and reduces energy cost from 1.55 to 1.05 (32.3% reduction), with 0% false positives.
- Dataset contains 7,161 images (1,556 annotated, 5,605 unannotated) with 11,151 weed instances annotated with bounding boxes and stem points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add stem coordinate regression to object detection | YOLOv7 · WSD | Dist: 2.9770 -> 2.4838 | Helped |
| Add stem coordinate regression to object detection | YOLOv7 · WSD | Accuracy: 75.37% -> 80.42% | Helped |
| Add stem coordinate regression to object detection | YOLOv7 · WSD | Cost: 1.55 -> 1.05 | Helped |
| Add object detection to stem regression (ablation) | YOLOv7 stem regression · WSD | Dist: 4.3306 -> 2.4838 | Helped |
| Add semi-supervised learning with unlabeled data (ablation) | YOLOv7 stem regression + detection · WSD | Dist: 2.4838 -> 2.1485 | Helped |
| Replace YOLOv7 with YOLOv8-M (ablation) | Stem detection model · WSD | Dist: 2.4838 -> 3.3219 | Hurt |
| Replace YOLOv7 with YOLOv10-M (ablation) | Stem detection model · WSD | Dist: 2.4838 -> 3.2691 | Hurt |
| Replace YOLOv7 with YOLOv7-X (ablation) | Stem detection model · WSD | Dist: 2.4838 -> 2.4426 | Helped |
| Lower the confidence threshold from 0.056 to 0.055 (ablation) | YOLOv7 inference · WSD inference | FP: 0.000 -> 0.005 | Hurt |
| Lower the confidence threshold from 0.055 to 0.050 (ablation) | YOLOv7 inference · WSD inference | FP: 0.005 -> 0.020 | Hurt |
Surprising findings
- Higher mAP does not imply better weed stem localization; YOLOv8 and YOLOv10 underperform YOLOv7 on stem distance despite better general detection benchmarks.
- Anchor-free YOLOv8 and YOLOv10 showed significant drift in stem predictions compared with anchor-based YOLOv7.
Limitations
- Evaluation uses a simulated real-world experiment rather than actual laser hardware deployment.
- Only three crops and one weed type (four species) from a single sensor and fixed height setup.
- Hyper-parameters were chosen empirically; the 0% FP result depends on a chosen classifier threshold.
Experimental setup
Datasets: WSD (Weed Stem Detection)
Benchmarks: WSD
Metrics: Euclidean distance (Dist) false positive rate weeding accuracy energy cost (laser shots) inference time mAP
Baselines: YOLOv7 vanilla detection YOLOv8 YOLOv10 Faster R-CNN SSD direct stem regression
Builds on: YOLOv7 mean teacher
Closest papers by meaning
- Combining Deep Learning and Street View Imagery to Map Smallholder Crop Types (AAAI 2024 · similarity 0.79)
- AgMMU: A Comprehensive Agricultural Multimodal Understanding Benchmark (NeurIPS 2025 · similarity 0.78)
- Insect-Foundation: A Foundation Model and Large-scale 1M Dataset for Visual Insect Understanding (CVPR 2024 · similarity 0.78)
- Deep Change Monitoring: A Hyperbolic Representative Learning Framework and a Dataset for Long-term Fine-grained Tree Change Detection (CVPR 2025 · similarity 0.78)
- OAM-TCD: A globally diverse dataset of high-resolution tree cover maps (NeurIPS 2024 · similarity 0.78)
- Depth-Aware Concealed Crop Detection in Dense Agricultural Scenes (CVPR 2024 · similarity 0.78)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.