← Back to project
WildTrack3D · Random training-data samples
RGB, manual 2D, pseudo 3D, patch depth and camera alignment
data inspectionseeded random sampleno model inference
Random samples, actual loader outputs
Twelve videos drawn before visualization, with no manual selection by appearance or label quality. These are training inputs and annotations, not model predictions.
| Cohort | Pool | Drawn |
|---|
| Current training manifest | 30 | 6 |
| Additional verified transfer pool | 10830 | 6 |
- Seed 20260917; 8 frames at 2 FPS, up to 16 initially visible objects. Photometric augmentation and depth dropout are disabled for inspection; spatial preprocessing is unchanged.
- Current-train clips preserve their manifest frame indices. Candidate clips use the training loader’s seeded random start. Training and held-out IDs are excluded from the candidate draw.
- 3D boxes and metric depth are pseudo-labels / estimates, not human 3D ground truth. Missing depth is shown explicitly; failed samples remain listed without replacement.
- Top row: source RGB, loader-aligned manual 2D, projected pseudo 3D. Bottom row: depth, RGB/depth overlay, valid-pixel coverage.
- Depth uses the same 0–20 m color range throughout. Values above 20 m saturate only in the display; downloaded tensors retain full FP32 values.
- Sampling provenance · Exact sampled manifest
Clip overview
- The final column counts object-frames with all 3D components valid, visible manual 2D and a measurement anchor. Partial-component supervision may exist elsewhere; this is not a total loss count.
sav_009516 · current train
- Current training manifest · object IDs [0, 1, 2, 3, 4, 5]
sav_003064 · current train
- Current training manifest · object IDs [0, 1, 2]
sav_010007 · current train
- Current training manifest · object IDs [0, 1, 2]
sav_010602 · current train
- Current training manifest · object IDs [0, 1, 2, 3]
sav_005705 · current train
- Current training manifest · object IDs [0, 1, 2]
sav_005633 · current train
- Current training manifest · object IDs [0, 1, 2, 3, 4, 5]
sav_010579 · candidate pool
- Additional candidate pool — not claimed as already trained · object IDs [0, 1, 2]
sav_004472 · candidate pool
- Additional candidate pool — not claimed as already trained · object IDs [1]
sav_000083 · candidate pool
- Additional candidate pool — not claimed as already trained · object IDs [0, 1, 2, 3, 4, 6, 7]
sav_002698 · candidate pool
- Additional candidate pool — not claimed as already trained · object IDs [0, 1]
sav_007078 · candidate pool
- Additional candidate pool — not claimed as already trained · object IDs [0, 1, 2]
sav_010482 · candidate pool
- Additional candidate pool — not claimed as already trained · object IDs [0, 2]