zuhdifr/sam3-mulchongsae

SAM3 fine-tune — 물총새_pcap / Mulchongsae Park (task11, segmentation)

Segmentation-head fine-tune of facebook/sam3, trained end-to-end (box + mask losses) on pseudo-RGB frames from a static roadside Ouster OS1-128 capture (물총새_pcap, cap__seg00-000.pcap), curated from CVAT task 11 annotations.

• Checkpoint: epoch 18 (checkpoint_best.pt), selected by best validation bbox AP.
• Validation: AP=0.598, AP50=0.934, AP75=0.609.
• Classes (SAM3 text prompts): person, bicycle, motorcycle, car, bus, truck.
• Format: flat SAM3Image state_dict, fully self-sufficient (own trained backbone + segmentation_head.*), not nested under a "detector." prefix like Meta's official release checkpoints — load with checkpoint_path=None then model.load_state_dict(state_dict, strict=False) manually, not via build_sam3_image_model's own checkpoint_path= loader.

Same training recipe as the sibling sam3-cap115414-task8-seg fine-tune (Sam3LossWrapper, matcher = BinaryHungarianMatcherV2 + o2m BinaryOneToManyMatcher; Masks loss point-sampled, final decoder stage only).

Part of a larger pipeline (2D SAM3 segmentation → box-based NMS → 3D box reconstruction → tracking-based refinement → SUSTechPOINTS export) for this LiDAR capture.