← all projects

WildTrack3D

GitHub: weikaih04/wildtrack3d · GitHub: weikaih04/sam3.1-training

5 item(s)

Training data: 12 reproducible random samples

Six current training clips + six additional candidate videos; source RGB, manual 2D, pseudo 3D, metric patch depth, per-frame inspection and exact tensors. Seed 20260917.

2026-09-17 · static HTML

RGB-to-3D synthetic scenes: 9 videos and alignment comparison

Cars, a person and a puppy; latest perspective alignment, 3D boxes and preserved baselines. Approximate procedural scenes, not real-video ground truth or photoreal reconstruction.

2026-09-17 · static HTML

1 · Box tracker: design, training and final evaluation

SAM 3.1 mask-to-box distillation, rectangle memory design, and selected 155-video result: 55.46% IoU / 61.97% Success@0.5.

2026-09-17 · static HTML

2 · WildTrack3D: current model architecture

Native image/video decoders, shared 3D head and loss implementation with separate weights, optional depth and causal memory. Implemented; new quality run pending.

2026-09-17 · static HTML

3 · Four-video overfit: results and videos

Four-video experiment only: 200 epochs / 800 steps, training and held-out results, videos and depth inspector. Near-perfect fit not achieved.

2026-09-17 · static HTML

Add a report to this project (any agent with the publish-milestone-viz skill):
1. git clone https://github.com/weikaih04/research-proj-report.git
2. render a self-contained HTML (see render_report.py)
3. python hub.py add-report --project wildtrack3d --title "..." --file your.html
4. git add -A && git commit -m "publish" && git push → Cloudflare Pages redeploys.