WildTrack3D · Complete 32-video box tracker evaluation

SA-V validation snapshot · September 15, 2026

2D box trackingSA-V validation32 video comparisonsLong-video scope

At a glance

The box-only tracker produces useful continuous tracking on SA-V validation clips, with visible success and failure cases available below.

Long-video robustness remains unverified. This report covers short clips and uses different input resolutions for the student and baseline.

Evaluation snapshot

System	Input	Visible box IoU	Success @ 0.5	Presence F1
Box tracker	1008 × 1008	57.78%	63.87%	89.59%
Mask-to-box baseline	672 × 672	61.27%	65.98%	86.64%
Student − baseline	Different resolutions	−3.49 pp	−2.12 pp	+2.94 pp

Custom box-tracking metrics, not official SA-V mask J&F.
Same videos, initial boxes, scored frames, and objects. This is a system comparison, not a same-resolution architecture ablation.
Snapshot: v3_1008_geomkd/student_1008.pt. Later tuning runs are outside this report.

What was actually tested

Coverage	Protocol
32 validation videos	First 32 video IDs in sorted order; all annotated frames, no frame cap
2.3–38.3 seconds	Clip duration; no minute-scale sequence in this cohort
6 fps propagation	Annotated frames from the 24 fps image release; real-time-speed playback
4,868 visible / 1,202 absent object-frames	Objects visible at initialization only; initialization frame excluded from scores
One initial box per object	No later corrections; predictions feed subsequent tracker memory

Success means IoU ≥ 0.5 on visible target frames, without presence gating. Presence F1 is reported separately.
Video-level IoU averages visible object-frames. The curves below average visible objects at each frame, so their unweighted mean can differ.
This validation cohort has been used during model development; results are not an untouched test-set estimate.

All 32 videos · index

Video	Seconds	Objects	Box IoU	Mask-to-box IoU
sav_000262	22.0	2	70.4%	32.4%
sav_001676	2.3	1	43.4%	28.9%
sav_001799	17.7	2	60.0%	71.8%
sav_001826	26.5	2	35.9%	38.1%
sav_001877	10.3	3	15.3%	22.3%
sav_001964	8.7	1	70.2%	79.2%
sav_002534	15.5	3	41.7%	50.2%
sav_003501	15.7	1	94.6%	99.0%
sav_004164	10.2	3	60.4%	84.3%
sav_004380	38.3	1	24.1%	25.6%
sav_004844	13.8	2	27.8%	52.6%
sav_004873	12.3	1	39.8%	39.9%
sav_005270	20.7	3	66.6%	59.2%
sav_005574	10.7	1	57.1%	95.3%
sav_005961	14.2	4	74.8%	89.4%
sav_006114	15.5	2	86.4%	89.9%
sav_006251	19.3	1	95.8%	98.8%
sav_006298	28.7	2	77.6%	85.3%
sav_006493	13.7	3	45.5%	40.4%
sav_006751	13.2	5	76.2%	79.4%
sav_006941	17.2	2	42.8%	47.7%
sav_009079	7.7	2	4.1%	51.6%
sav_009503	28.2	1	55.5%	77.1%
sav_009628	14.2	1	29.5%	36.9%
sav_010291	29.3	1	71.1%	35.3%
sav_010406	12.8	1	31.1%	27.5%
sav_010810	17.3	2	54.1%	61.5%
sav_010922	11.2	1	59.7%	45.0%
sav_011301	32.0	2	45.5%	57.0%
sav_011413	14.3	1	43.5%	75.7%
sav_011732	22.5	2	53.2%	57.8%
sav_011944	19.2	1	84.4%	68.3%

1 · sav_000262

Duration	Initial objects	Box IoU	Mask-to-box IoU
22.0 s	2	70.4%	32.4%

Full evaluation cohort. Left: box tracker at 1008; right: mask-to-box at 672. Dashed boxes: ground truth; solid boxes: predictions; colors: object IDs; p: presence probability. Low-confidence predictions remain visible.