zuhdifr/sam3-cheonan-sicheong

SAM3 fine-tune — cap_20260716_115414 (task8, segmentation)

Segmentation-head fine-tune of facebook/sam3, trained end-to-end (box + mask losses) on 450 train / 50 val pseudo-RGB frames from a static roadside Ouster OS1-128 capture (cap_20260716_115414, 6107 frames @ 10Hz), curated from CVAT task 8 annotations.

• Checkpoint: epoch 18 (of 20), selected by best validation bbox AP.
• Validation: AP=0.573, AP50=0.774, AP75=0.690.
• Classes (SAM3 text prompts): person, bicycle, motorcycle, car, bus, truck.
• Format: flat SAM3Image state_dict (1132 keys, incl. backbone + trained 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.

Loss (Sam3LossWrapper, matcher = BinaryHungarianMatcherV2 + o2m BinaryOneToManyMatcher): Boxes loss_bbox=5.0/loss_giou=2.0; IABCEMdetr loss_ce=20.0/presence_loss=20.0/pos_weight=10/focal α=0.25,γ=2; Masks loss_mask=200.0/loss_dice=10.0, point-sampled (12544 points, oversample_ratio=3.0, importance_sample_ratio=0.75), final decoder stage only.

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