MotionForesight benchmark: SSv2 held-out 150 + OOD phone videos 50

Evaluation data and scoring code for MotionForesight (arXiv 2607.16192, Tables 1 and 2), with the released checkpoint's predictions for checking your setup.

The SSv2 clips are derived from Something-Something v2; please do not redistribute them.

ssv2_val150/
  clip_ids.txt, manifest.json
  clips/<vid>.npz           one clip (keys below)
  masks/<vid>.png           frame-0 object mask (query points were sampled inside it)
  videos/<vid>.mp4          22-frame preview, for viewing only (use clips/*.npz rgb)
ood_phone50/                 phone videos, uniform sampling (Table 1)
  clip_ids.txt, manifest.json, clips/, videos/
  full_videos/<vid>.mp4     the complete phone recordings
ood_phone50_longcontext/     the same phone videos, long-context sampling (Table 2)
  clip_ids.txt, manifest.json, clips/, videos/
mf_bench.py                  loader + ADE/FDE/PWT scorer (numpy only)
reference_predictions/motionforesight_40k/
  <split>/<vid>.npz          released-model predictions, pred_tracks_3d (22,P,3), last-observed camera
  <split>_results.json       their scores
tools/preprocess_ood_videos_orient.py   Depth Anything 3 preprocessing used for the phone videos

Protocol

Every clip has 22 frames. Observe frames 0–6, forecast the 3D position of every frame-0 query point (query_xy) at frames 7–21.

• SSv2: 22 consecutive frames at 12 fps.
• OOD, uniform: round(linspace(0, N-1, 22)) over the N decoded frames of the recording (N = 74–188, ~30 fps).
• OOD, long context: mid = max(7, N // 2); observed = round(linspace(0, mid-1, 7)), future = round(linspace(mid, N-1, 15)).
• source_frame_idx and timestamps_s give the exact frames for every clip.

MotionForesight's inputs are rgb[0:7], depth and cameras for frames 0–6, and the frame-0 query points. Future RGB, depth and cameras are for scoring only. action is a text description of the clip and is not an input to MotionForesight.

clips/<vid>.npz

key	shape / type	meaning
rgb	(22,480,832,3) uint8	the frames
obs_frames, total_frames	int	7, 22
observed_frame_idx, future_frame_idx	(7,), (15,)	0–6 and 7–21
source_video_id	str	SSv2 video id / phone video name
source_frame_idx	(22,) int	frame index in the source video
source_num_frames	int	OOD only: decoded frame count of the full recording
source_fps, timestamps_s	float, (22,)	source fps, and source_frame_idx / fps
query_xy	(P,2) float32	frame-0 query pixels (x, y) at 480×832
tracks_3d_lastobs	(22,P,3) float32	GT tracks, metres, camera frame of frame 6 (the scoring frame)
tracks_3d_world	(22,P,3) float32	the same tracks in the frame-0 camera frame
valid	(22,P) bool	scoring mask
w2c	(22,4,4) float32	frame-0 camera → camera t (w2c[0] = I). OpenCV axes: x right, y down, z forward
intr	(4,)	fx, fy, cx, cy at 480×832
depth	(22,h,w) float32	metric depth (Depth Anything 3) at depth_hw; intr_depth are the intrinsics at that size. Not in the long-context split
action	str	text description of the clip

tracks_3d_lastobs[t] = R6 · tracks_3d_world[t] + t6, with [R6|t6] = w2c[6].

Scoring

python mf_bench.py check --data-dir ssv2_val150
python mf_bench.py score --data-dir ssv2_val150 --pred-dir <your_preds> [--frame lastobs|world]

Prediction files are <pred_dir>/<vid>.npz with pred_tracks_3d of shape (22,P,3) or (15,P,3), in metres, queries in query_xy order.

• ADE: per clip, mean L2 over valid (future frame, query) pairs; then mean over clips (cm).
• FDE: per clip, mean L2 over valid queries at frame 21; then mean over clips (cm).
• PWT: fraction of valid future errors within τ, pooled over all clips, frames and queries, for τ ∈ {1, 2, 5, 10, 20} cm, averaged over the five thresholds. PWT@5cm is printed too.
• No alignment or scale fitting.

Released checkpoint (trainedon_ss100k/best.pt):

split		ADE (cm)	FDE (cm)	PWT (%)	PWT@5cm (%)
SSv2 150	paper, Table 1	4.47	6.23	76	—
	reference_predictions/	4.47	6.24	76.0	82.5
OOD 50, uniform	paper, Table 1	9.31	14.88	54	—
	reference_predictions/	9.31	14.88	54.3	53.6
OOD 50, long context	paper, Table 2	10.20	13.59	—	47.4
	reference_predictions/	10.20	13.59	48.8	47.4

Table 2 reports PWT@5cm. Inference runs in bf16, so reruns can move the aggregates by about ±0.01 cm.

Notes

• GT tracks come from TrackCraft3r run on the 22 frames with Depth Anything 3 depth and cameras, read at the query pixels. There is no sensor ground truth. valid is all True.
• Query points: an 8-pixel grid inside the frame-0 object mask, capped at 2000 per clip.
• The phone videos carry 180° rotation metadata; rgb, cameras and GT use the upright decode. If you decode full_videos/ yourself, apply it (OpenCV: cap.set(cv2.CAP_PROP_ORIENTATION_AUTO, 1)).
• Tracks exist only for the 22 frames of each clip. The full phone recordings are included for longer contexts and horizons; SSv2 source videos are not.
• ood_phone50_longcontext has no depth. To regenerate depth and cameras, run tools/preprocess_ood_videos_orient.py --sampling split --split-obs 7 --num-frames 22 on ood_phone50/full_videos/. Depth Anything 3 camera estimates vary slightly between runs under this sampling, so regenerated cameras will not match the stored ones exactly; score against the stored GT.