Spatial Encoder Evaluation Dataset (SEED)

Short video clips with keyframe camera poses and 3D oriented bounding boxes (OBBs) in a per-clip world frame, for evaluating open-vocabulary 3D object detection from monocular video without camera pose input. Two releases live in this repo, each with its own README, tables and WebDataset shards:

Release	Folder	Clips	Keyframes	3D boxes	Shards	Annotation
SEED v2	SEED_v2/	334	15,008	46,468	7 (5.9 GB)	Human-annotated ground truth (KITTI-360, NymeriaPlus, CODa, corrected Boxer)
SEED v1	SEED_v1/	923	42,096	37,983	19 (18.8 GB)	Mostly model pre-annotations with LLM review (SpatialVID, Sekai) + Boxer + CODa

Both releases share the clip format, the 314-class vocabulary (vocab314.csv), the per-clip positive / negative category sets for federated scoring, and the metric (tools/eval_metric.py). Use SEED v2 for reporting numbers; v1 is kept for reproducibility of earlier results. Nothing in v1 was modified; its files were moved from the repo root to SEED_v1/.

Download

pip install -U huggingface_hub
hf download wangchenlong/SEED --repo-type dataset --include "SEED_v2/*" --local-dir seed   # v2 only (5.9 GB)
hf download wangchenlong/SEED --repo-type dataset --include "SEED_v1/*" --local-dir seed   # v1 only (18.8 GB)

Then follow SEED_v2/README.md (or SEED_v1/README.md) for unpacking the shards and running the metric.

Licenses

Non-commercial research use only. SEED v2 is released under CC BY-NC-SA 4.0 (KITTI-360 clips: CC BY-NC-SA 3.0; NymeriaPlus clips: CC BY-NC 4.0; Boxer / CODa clips: as in SEED v1). SEED v1 is released under CC BY-NC 4.0. Please cite the source datasets as listed in each release's README.