wheat-point-detect

wheat-point-detect — DINOv3 root-point detector (area-aux POC, epoch 200)

Frozen DINOv3 ViT-B/16 backbone + dual-branch decoder (density / classification / regression / polygon-area auxiliary head, area_lambda=0.15). The checkpoint is run 0003-point18-joint-area-poc of the joint-root-area-poc experiment: 200 epochs, seed 1001, patch 512 / stride 512, AdamW lr 1e-4, batch size 2 (2x1 microbatch with gradient accumulation, fp16 autocast), trained on one 3797x3943 wheat tile with 209 manual CVAT root points plus SAM3 pseudo-area supervision.

Files

Operating point

RGB input, white pixels (all channels >= 250) filled black, ImageNet normalization, 512/512 tiled accumulation with averaging; points = peak_local_max(sigmoid(classification), threshold_abs=0.5, min_distance=8) plus regression offsets.

Same-tile full image: P 0.801, R 0.636, F1 0.709 (tp 133 / fp 33 / fn 76, count error 43 of 209 roots). Held-out spatial region (same tile, y>=2048): F1 0.548. Matching tolerance 15 px.

Scope and limitations (from RUN_MANIFEST)

This is a single-tile architecture/behavior POC, not an area-supervision effect claim: no matched point-only baseline, spatial holdout from the same source tile instead of unseen images, and the area target is SAM3 pseudo segmentation rather than manual polygons. The unseen-image predictions (166 points) are qualitative only.

Serving

A self-contained Nuclio function kit (Dockerfile, function-gpu.yaml, handler) lives in nuclio/point-detector/ of Samuka007/cits5551-group7 (branch nuclio). The backbone snapshot must be fetched separately from the gated repo facebook/dinov3-vitb16-pretrain-lvd1689m.