hmarchant/countgd-litter-dense

CountGD++ Litter — dense fine-tune

CountGD++ fine-tuned to count litter in photos, including extremely dense illegal-dumping piles.

File	What it is
countgd_litter_noleaf.pth (recommended)	Trained on labels with fallen leaves and other plant matter removed
countgd_litter_dense.pth	Same recipe on the uncleaned labels; counts fallen leaves as litter

• Prompt: litter · recommended confidence 0.23
• Backbone/BERT frozen; detection head fine-tuned with token focal loss + L1/GIoU on SAM 3 pseudo-labels
• Teacher: Meta SAM 3 (facebook/sam3), four-prompt union (litter, trash, garbage, discarded packaging), with 4×4 tiled inference on human-flagged extreme-density images
• Training set: 954 human-triaged images (194 extreme-density)
• Leaf clean-up: Claude Haiku 5.5 reviewed every teacher box and removed 3,158 of 35,529 (8.9%) that sit on leaves, plants, stones or soil; a box is removed only when two independent reviews agree

Results

	litter_dense	litter_noleaf
Holdout count MAE vs cleaned labels (200 images)	11.7	10.8
Holdout predictions on removed leaf boxes	324	234
48-image comparison set, mean count	65.8	60.0
Street view (119 KartaView images): detections judged natural by Haiku	591 (43%)	334 (31%)
Street view: detections judged litter by Haiku	324	377
Leaf-covered cobbled square, Germany	106	0

Checkpoint format

torch.load(path, map_location="cpu", weights_only=False) → {"model": state_dict, "epoch": 2, "prompt": "litter", "holdout": [...]}. Load ckpt["model"] into CountGD++ (countgd_infer.py in https://github.com/HugoMarchant/llitter-index).

Limits

Counts are pseudo-label quality: SAM 3 still undercounts the very densest piles. Leaves remain the largest false-positive class (31% of street-view detections). Counts are capped by CountGD++'s query budget (~900). Trained on litter scenes; not a general object counter.

See RESULTS.md in the code repository for the full report.