Dataset Viewer (First 5GB)
Auto-converted to Parquet Duplicate
The dataset viewer is not available for this split.
Parquet error: Scan size limit exceeded: attempted to read 758473967 bytes, limit is 300000000 bytes Make sure that 1. the Parquet files contain a page index to enable random access without loading entire row groups2. otherwise use smaller row-group sizes when serializing the Parquet files
Error code:   TooBigContentError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

JtT Kubric 2D/3D dense point tracks

Preprocessed Kubric MOVi-A and MOVi-D videos with dense 2D and 3D point tracks, depth, segmentation and camera parameters, as used to train the JtT track generator, together with the two multi-rollout benchmarks it is evaluated on.

Each video is 24 frames at 256x256. For every video, 4096 points are sampled on query frame 0 and tracked through all frames in both image space and 3D camera space.

Contents

Subset Videos Tar size Parts
movi_a_2d_3d 8144 train, 211 validation 63 GB part00 to part05
movi_d_2d_3d 7906 train, 197 validation 72 GB part00 to part06
movi_a_2d_3d_benchmark 1024 (32 scenes x 32 rollouts) 16 GB part00 to part01
movi_d_2d_3d_benchmark 1024 (32 scenes x 32 rollouts) 19 GB part00 to part01

Each subset is one uncompressed tar, split into 10 GB parts so that every file stays under the Hub's 50 GB limit.

movi_a_2d_3d/movi_a_2d_3d.tar.part00 ... part05
movi_d_2d_3d/movi_d_2d_3d.tar.part00 ... part06
movi_a_2d_3d_benchmark/movi_a_2d_3d_benchmark.tar.part00 ... part01
movi_d_2d_3d_benchmark/movi_d_2d_3d_benchmark.tar.part00 ... part01

Download and extract

hf download Gabrijel/jtt-kubric --repo-type dataset --local-dir jtt-kubric
cd jtt-kubric
mkdir -p movi_a_2d_3d_extracted movi_d_2d_3d_extracted
cat movi_a_2d_3d/movi_a_2d_3d.tar.part* | tar xf - -C movi_a_2d_3d_extracted --strip-components=1
cat movi_d_2d_3d/movi_d_2d_3d.tar.part* | tar xf - -C movi_d_2d_3d_extracted --strip-components=1
cat movi_a_2d_3d_benchmark/movi_a_2d_3d_benchmark.tar.part* | tar xf - -C movi_a_2d_3d_extracted --strip-components=1
cat movi_d_2d_3d_benchmark/movi_d_2d_3d_benchmark.tar.part* | tar xf - -C movi_d_2d_3d_extracted --strip-components=1

To download only the benchmarks, add --include "*_benchmark/*" to hf download.

This produces, for each of movi_a_2d_3d_extracted and movi_d_2d_3d_extracted:

train/query_frame-00/*.h5
validation/query_frame-00/*.h5
benchmark_subsampled_renamed/query_frame-00/scene_XX_rollout_YY.h5
benchmark_subsampled_renamed_trackbench/scene_XX_rollout_YY/

Without --strip-components=1 the tars unpack into their original folder names, kubric_movi_a_dense_2d_3d_preprocessed/ and kubric_movi_d_dense_2d_3d_preprocessed/.

File format

One HDF5 file per video (hdf5plugin is required to read the compressed datasets).

Dataset Shape Dtype Description
video (24, 256, 256, 3) uint8 RGB frames
depth (24, 256, 256) float32 Metric depth
inv_depth (24, 256, 256) float32 Inverse depth
segmentation (24, 256, 256, 1) uint8 Instance segmentation
tracks_2d (4096, 24, 2) float32 Pixel coordinates (x, y) per point per frame
tracks_3d (4096, 24, 3) float32 3D coordinates per point per frame
occluded (4096, 24) bool Occlusion flag
valid_mask (4096, 24, 3) bool Validity of each track coordinate
patch_2d_centers (256, 2) float32 Centers of the 16x16 query patches
patch_3d_coords (256, 3) float32 3D coordinates of patch centers
camera__positions (24, 3) float32 Camera position per frame
camera__quaternions (24, 4) float32 Camera orientation per frame
camera__focal_length, camera__sensor_width, camera__field_of_view () float32 Intrinsics
video_name (1,) str Source Kubric video id

The 4096 points form a 64x64 grid on frame 0 with a stride of 4 pixels (pixel centres 0.5, 4.5, ..., 252.5), flattened row by row.

import hdf5plugin, h5py
with h5py.File("movi_a_2d_3d_extracted/train/query_frame-00/1.h5") as f:
    video, tracks_2d = f["video"][:], f["tracks_2d"][:]

Multi-rollout benchmarks

Each benchmark contains 32 scenes with 32 rollouts per scene. All rollouts of a scene start from the same first frame and are simulated with different rollout seeds, so they diverge afterwards. A scene therefore provides 32 ground-truth futures for one input frame, and a generative model is evaluated by comparing its samples for that frame with the set of 32 simulated futures.

  • MOVi-A: 32 scenes with 3 to 10 objects (2 to 5 scenes per object count), 16 from each of the Kubric master seeds 42 and 7919.
  • MOVi-D: 32 scenes from Kubric master seed 42, chosen at random among scenes with at least 32 rollouts left after preprocessing, with 32 rollouts chosen at random per scene.

Scenes are numbered scene_01 to scene_32 and rollouts rollout_01 to rollout_32. Every rollout is stored twice:

benchmark_subsampled_renamed/query_frame-00/scene_XX_rollout_YY.h5: the same HDF5 format as the training files above (model input), with two differences: depth, inv_depth and patch_3d_coords are stored as float64, and the MOVi-A files have no segmentation (the masks are in the export below).

benchmark_subsampled_renamed_trackbench/scene_XX_rollout_YY/: the same rollout exported for evaluation:

File Content
00.png ... 23.png RGB frames, identical to video
00_mask.png ... 23_mask.png Instance segmentation (uint8, 256x256)
tracks_00.npz tracks (24, 64, 64, 2) float32, pixel (x, y)
vis (24, 64, 64) bool, True for occluded points (same as occluded)
tracks_3d (24, 64, 64, 3) float32; world_pos is a copy of tracks_3d

The arrays in tracks_00.npz are tracks_2d, occluded and tracks_3d from the HDF5 file, reshaped from 4096 points to the 64x64 grid and moved to time-first order. For evaluation we use every other grid point, tracks[:, 1::2, 1::2], i.e. 1024 tracks on a 32x32 grid with a stride of 8 pixels.

Source

Generated from Kubric MOVi-A and MOVi-D, which are released under the Apache 2.0 license.

Downloads last month
127