hmarchant/sam3-litter-sv

SAM 3 street-view litter counter

SAM 3 (image model) fine-tuned to find litter in street-view photos with the single prompt litter. It counts litter for a litter index that compares countries and continents.

• Checkpoint: sam3_litter_sv.pt (weights under detector. keys, as the released SAM 3 checkpoint).
• Input: 1,008 px square (native), image not shrunk beforehand.
• Prompt litter, confidence 0.58 (presence × box score). A count is the number of boxes above it.
• Training: 1,648 Haiku-labelled Mapillary street-view images, with unsure items as negatives at weight 0.5 and mined not-litter boxes as negatives at weight 2; images with 10+ items drawn four times as often. Full fine-tune, epoch 4 of 6, chosen on a 1,656-image validation set.
• Test (1,536 fresh Mapillary images, Haiku-built reference): count MAE 0.760, bias +0.07 per image, recall 48%, precision 45%; every continent's bias within ±max(0.3, 25% of its reference mean).

Run with scripts/sam3_infer.py from the litter-index repository, inside a clone of facebookresearch/sam3:

python sam3_infer.py --input IMAGES --output OUT --checkpoint sam3_litter_sv.pt --resolution 1008 --max-side 0 \
  --text litter --conf 0.58

The weights derive from SAM 3 and are subject to the SAM License.