sahilchat/static-scene-test2

DL3DV 480p H.264 scene containers

DL3DV-10K scenes packed for streaming novel-view-synthesis training with spatial_rep.data.streaming.hf_scenes (pairlab/spatial-rep-learning, branch sahil/hf-indexed-streaming). Every GOP is independently fetchable by HTTP byte range and decodable.

Source and resolution

• Source: DL3DV/DL3DV-ALL-960P (images_4 PNG frames at 960x540 plus nerfstudio transforms.json, camera model OPENCV).
• Stored at 852x480 (frames downscaled with a box filter to height 480, even width; the ~4% of scenes whose 960P zip only holds 480x270 frames are stored at that size, never upscaled), undistorted with the OPENCV k1 k2 p1 p2 coefficients at pack time, so the stored camera model is PINHOLE.
• Poses converted from the OpenGL camera convention of transforms.json to OpenCV axes (+X right, +Y down, +Z forward); intrinsics normalized by width and height.
• Scene lists: the joint-NVS DL3DV split (10,013 train / 130 val scene hashes).

Encoding

libx264, CRF 18, preset medium, GOP 8 (keyint=min-keyint=8, scenecut=0), closed GOPs, no B-frames, yuv420p, SPS/PPS repeated at every IDR frame (repeat-headers=1, Annex-B).

Measured with pack_scenes rate on 4 validation scenes, first 64 frames each at 852x480, GOP 8:

CRF	PSNR (dB)	SSIM	KB/frame	KB/keyframe	vs raw RGB
16	37.1	0.979	47.5	84.3	34x
18	36.2	0.973	38.0	71.2	43x
20	35.2	0.964	30.1	59.6	56x
23	33.6	0.946	20.7	44.9	81x
28	31.0	0.898	10.7	26.5	153x

Container layout

One HDF5 file per shard of up to 48 scenes, train/hdf5/shard-TT-XXXX.h5 and val/hdf5/..., indexed by train/manifest.json and val/manifest.json; index/<shard>.npz holds each shard's tables and blob chunk map so the loader indexes a shard with one small request. Datasets per shard (N frames over S scenes): blob uint8 [B] (all access units back to back), offsets int64 [N+1], keyframe uint8 [N], scene_offsets int64 [S+1], c2w float32 [N,4,4], intr float32 [N,3,3], split uint8 [N] (0 train view, 2 held-out target), camera_id int32 [N], scenes_json str. See the module docstring of spatial_rep/data/streaming/pack_scenes.py for the authoritative description and the packer that produced these files.

Contents

split	scenes	frames	shards	size
train	10,013	3,421,619	224	109.2 GB
val	130	45,676	3	1.5 GB

Total 110.7 GB, 3,467,295 frames, 31.2 KB per frame on average.

Streaming from the Hub

Measured 2026-10-20 from one compute node with an authenticated token. The CDN that serves file bytes is not rate limited: 3,000 range GETs of 256 KB at 174 requests/s and 150 whole-file GETs (20 GB at 830 MB/s) returned no errors. huggingface.co meters GET requests on resolve/ URLs at 5,000 per account per fixed 5-minute window (RateLimit-Policy: "fixed window"; "resolvers"; q=5000; w=300; over quota it answers 429 with Retry-After of about 3 minutes). HEAD requests on the same URLs are not counted (6,000 in 37 s, no 429). spatial_rep.data.streaming.hf_scenes.HFByteSource resolves each file once with a HEAD and then reads the signed CDN URL, so a training job makes no metered requests; a client that GETs a resolve/ URL per file (whole-tar streaming) spends one metered request per file per epoch.

Licence

Derived from DL3DV-10K, released under CC BY-NC 4.0: non-commercial research use only, attribution to the DL3DV authors required. Please cite the DL3DV-10K paper (Ling et al., CVPR 2024) when using this data.