vveicao/infinigen-traj-2k

Infinigen Trajectory Videos 2K

2,000 synthetic indoor videos rendered from Infinigen scenes, with per-frame RGB, metric depth, camera intrinsics and poses, instance masks and per-frame 3D oriented boxes. Each video is a slow, wall-avoiding camera path between two real camera poses of one scene. The data is in the native scene-JSON format of SpatialEncoder v2 Small (spatial_encoder_v2_small), plus WebDataset shards that its HF/WDS training loader can stream.

	train	val	total
Videos	1,822	178	2,000
Frames	179,400	17,580	196,980
Object tracks			39,303 (73 categories)
WDS shards / samples	74 / 3,609	8 / 353	81.5 GB

• 1,003 layouts of about 1,000 Infinigen scenes, 1 or 2 videos per layout; the split follows the layout's split.
• 45–120 frames per video (multiples of 15, mean 98.5), 15 fps, 1024 × 1024.
• Camera: 0.3 m/s, 20°/s, path length 0.8–2 m, total turn ≤ 90°; about 2 cm and 0.5° of motion per frame, similar to ScanNet (0.9 cm / 0.7° median).
• Horizontal field of view fixed per video, 48–82° across videos (median 67°); pinhole, no distortion.

Layout

README.md
manifest.json                         WDS shard manifest (schema spatialencoder-wds-native-v1, pinned URLs)
catalog.json                          training catalog with the "Infinigen" source ($INFINIGEN_ROOT paths)
dataset/Infinigen.json                index: [["train-json/<video>.json", num_frames, "train"|"val"], ...]
Infinigen/train-json/<video>.json     one scene JSON per video (all splits; the index gives the split)
videos/<video>.tar                    per-video package, extracts to <video>/...
  <video>/render/rgb_frames/NNNNNN.jpg   RGB
  <video>/render/rgb.mp4                 the same frames as a video
  <video>/depth/depth_mm_uint16/NNNNNN.png   depth, uint16 millimetres, 0 = invalid
  <video>/depth/depth_preview.mp4
  <video>/instance/instance_id_png/NNNNNN.png   instance ids
  <video>/annotations/                   renderer output: video_annotation.json, per_frame/*.json, categories, source
  <video>/trajectory.json                planned camera path
wds/Infinigen/infinigen-<split>-NNNNNN.tar   ~1 GB WebDataset shards (rgb + depth + metadata)
metadata/registry.jsonl               every scene, layout and Part 3 camera used (for de-duplicating later batches)
metadata/selections/                  layout selection lists
metadata/trajectory_plans.tar         per-layout trajectory plans
metadata/stats.json                   counts and the category histogram
metadata/videos_sha256.json           size and sha256 of every videos/*.tar

Conventions

Scene JSON frames[frame_id]:

Field	Meaning
rgb_path, depth_path, instance_path	relative to videos/ (after extracting the tars)
rgb_intrinsic, depth_intrinsic	3 × 3 K in pixels, pixel centres at integer coordinates (cx = cy = 511.5)
camera_pose	4 × 4 camera-to-world, metres, OpenCV camera axes (+X right, +Y down, +Z forward), world +Z up
original_width, original_height	1024, 1024
time_s	frame time at 15 fps

Scene JSON objects[object_id][frame_id] (only frames where the object is visible):

Field	Meaning
camera_center	box centre in the camera frame, metres
camera_scale	full side lengths, metres, in the order of camera_R's columns
camera_R	box-local → camera rotation; local Y is the box axis closest to world vertical, pointing down
camera_box_2d_rend	2D prompt / 2D target used by the loader: envelope of the projected cuboid, clipped to the image [-0.5, W-0.5] × [-0.5, H-0.5] (same convention as ScanNet val-json and ADT)
camera_box_2d_visible	box of the visible instance mask
camera_box_2d_proj	envelope of the projected cuboid, not clipped
category_name, visible_pixels	query text and visible pixel count

Object-level fields: 3d_category (query text, Part 3 shared vocabulary with _ → space), part3_category, instance_id, is_added (object added by Infinigen layout augmentation). Doors and windows are not exported as objects.

Depth is planar z in millimetres (default_scale 1000).

Using it with spatial_encoder_v2_small

The training code builds an ObjectDetectionDataset only for names in DATASET_IDS (sam3/train/data/object_detection_dataset.py), so register the source first, e.g. DATASET_IDS.setdefault("Infinigen", max(DATASET_IDS.values()) + 1) in every training process.

Native files.

huggingface-cli download vveicao/infinigen-traj-2k --repo-type dataset --local-dir $INFINIGEN_ROOT \
  --include "videos/*" "Infinigen/*" "dataset/*" "catalog.json"
for f in $INFINIGEN_ROOT/videos/*.tar; do tar -xf "$f" -C $INFINIGEN_ROOT/videos && rm "$f"; done
export INFINIGEN_ROOT
# then train with ++trainer.data.train.dataset.catalog_path=$INFINIGEN_ROOT/catalog.json

WebDataset streaming. manifest.json follows the same schema as the xxxspatialencoderwds*/data_* releases, so training/prepare_hf_wds.py can read this repository as one more manifest repo. One sample is a block of 64 clip starts plus 14 look-ahead frames of one video (the native loader's 15-frame windows); shards mix videos. To add the source to a lock, include vveicao/infinigen-traj-2k in the repos passed to prepare_lock, add "Infinigen" to EXPECTED_SOURCES (sam3/train/data/wds_manifest.py) and to wds.train_sources / wds.val_sources, and set its weight in wds.train_weights.

How it was made

• Scenes: Infinigen indoor .blend files (Part 1), with camera poses, categories and splits from the matching Part 3 release; every .blend was checked against the Part 3 sha256 (2,000 / 2,000).
• Paths: Dijkstra on a 0.1 m floor grid, at least 0.2 m from walls and objects, then straightened and smoothed; height, yaw and pitch interpolated linearly, roll 0; the first and last frames equal the Part 3 camera poses. Views keep median depth ≥ 0.8 m and at least 2 trainable objects in ≥ 90 % of frames.
• Rendering: Blender 5.0.1 Cycles, 64 samples + OIDN denoising for RGB; a separate geometry pass for depth and instance ids.
• Categories: the Part 3 name of the same scene object (73 categories here); objects Part 3 does not know (mostly architectural pillars) keep the renderer's name.

Quality checks

• Depth reprojection between adjacent frames (K, poses, depth): ≥ 98.45 % of overlapping pixels agree within 1 % (median 99.9 %).
• The spatial_encoder_v2_small training loader reads boxes that match the JSON (< 1e-6 m, 0.02°) on sampled clips.
• WDS shards decode with the repository's own TarCursor and SpatialWDSDecoder; sample counts match the manifest.

Differences from real indoor scans

Compared with ScanNet val88: objects are farther from the camera (median 2.68 m vs 2.12 m), cameras are slightly higher (median 1.62 m vs 1.47 m above the floor), the field of view is wider (median 67° vs 58°), and the category mix differs (chair 3 % of tracks vs 30 %). In a fine-tuning test, a model trained only on this data shifted its distance prior and placed ScanNet objects 15–25 % farther; mixing it with real data is recommended.

Checkpoints

checkpoints/infinigen_10ep_lr1e-5/ holds SpatialEncoder v2 Small (Lite) fine-tuned on this dataset only (train split), starting from the e100 checkpoint (SpatialCoding/Data_verification, se_small_lite_wds518_e100/checkpoint.pt; these weights derive from it and follow its terms).

• 10 epochs × 1,600 clips per GPU on 8 GPUs (16,000 steps), batch 1 clip per GPU, 518 px, AdamW, learning rate 1e-5 → 1e-7 cosine (spatial-encoder group 1e-7 → 1e-9), same losses and augmentation as e100.
• checkpoint_epoch{2,5,10}.pt are full trainer checkpoints (model, optimizer, epoch, ...), 2.4 GB each; config.yaml is the resolved training config; val88_metrics.json has the scores below.
• Trained on the first version of the scene JSONs, where camera_box_2d_rend was the visible-mask box (now camera_box_2d_visible).

ScanNet val88 (GT 2D-box prompts):

Checkpoint	[email protected]	[email protected]	[email protected]	3D IoU
e100 (start, not in this repo)	51.80	8.82	78.10	0.258
epoch 2	40.20	4.98	71.98	0.222
epoch 5	42.24	6.13	71.41	0.225
epoch 10	41.99	5.64	70.92	0.223

The drop comes mainly from the model's distance prior shifting toward this dataset's farther objects (see above); the same checkpoints improve sharply on this dataset's val split (rotation error 32° → 14°, centre error 0.67 → 0.41 m).