Grass Quality Segmentation
Includes videoA grass quality segmentation system that converts weak labels into patch-level supervision and produces quality heatmaps.
Project media
See the system in action.
Third-party embeds include a direct source link if browser privacy settings block playback.
Project Snapshot
| Signal | Details |
|---|---|
| Timeline | Mar 2025 to Apr 2025 |
| Organization | Benign Innovations |
| Focus | Grass Quality Segmentation |
| Stack | SAMv2, OpenCLIP, CLIP, SEEM, ResNet-18 |
Data Pipeline
- Generated SEEM + SAMv2 labels for dirt and grass classes.
- Converted YOLO labels into a grass-quality format: dirt and winter grass become bad, summer grass becomes good.
- Used CLIP to filter mislabeled data and create
64x64patches. - Applied OpenCLIP scoring to keep patches as good unless confidence is
>= 85%for bad, and vice versa.
Model
- Trained a ResNet-18 patch classifier with
64x64input. - Ran stride-1 inference to produce grass quality heatmaps.
Demonstration
The embedded demonstration shows the resulting grass-quality predictions and spatial heatmap output.