ZhengGuangze/PointOdyssey_vlbm

PointOdyssey (converted to VLBM format)

This dataset contains 131 training sequences from the PointOdyssey dataset converted to the VLBM-compatible format using preprocess_pointodyssey.py. The sequences have been compressed into .tar.gz archives in chunks of 50 sequences per archive.

Scale

Metric	Value
Total sequences	131
Image resolution	960 x 540 px
Depth type	Dense (ground truth)

Dataset Structure

Each sequence directory follows this layout:

<seq_name>/
├── rgbs/
│   ├── rgb_00000.jpg
│   ├── rgb_00001.jpg
│   └── ...
├── depths/
│   ├── depth_00000.npz
│   ├── depth_00001.npz
│   └── ...
├── intrinsics.npy
├── extrinsics.npy
├── trajs_2d.npy
├── trajs_3d.npy
├── visibilities.npy
└── scene_info.json

File Descriptions

• rgbs/: RGB frames saved as JPEG (rgb_XXXXX.jpg). Resolution is 960x540 pixels.
• depths/: Dense depth maps saved as compressed NumPy archives (depth_XXXXX.npz). Each archive stores a float16 array under the key depth of shape (H, W) in meters. Original PointOdyssey depth (uint16 PNG, max uint16 maps to 1000m) converted to float16 meters.
• intrinsics.npy: Camera intrinsic matrices for each frame (T, 3, 3) float16.
• extrinsics.npy: World-to-camera extrinsic matrices (W2C) for each frame (T, 4, 4) float16.
• trajs_2d.npy: 2D trajectories (T, N, 2) float16 -- pixel coordinates (x, y).
• trajs_3d.npy: 3D trajectories (T, N, 3) float16 -- world-space coordinates (x, y, z).
• visibilities.npy: Visibility flags (T, N) float16 (1.0 visible, 0.0 not visible).
• scene_info.json: JSON file with per-sequence metadata including num_frames, image_size, num_trajectories, source, and depth_range.

Trajectory Source

PointOdyssey provides ground-truth dense point trajectories in anno.npz files, containing trajs_3d, trajs_2d, visibs, intrinsics, and extrinsics. The conversion maps visibs to visibilities and tiles single intrinsic/extrinsic matrices across all frames when needed.

Data Specifications

• Image format: JPEG (RGB), 960x540 px
• Depth format: NPZ (float16), dense (ground truth from PointOdyssey)
• Annotation format: Individual .npy files (float16)
• Extrinsics: World-to-camera (W2C) 4x4 matrices

Usage Example (Python)

import numpy as np
from PIL import Image
from pathlib import Path
import json

seq_dir = Path("data/pointodyssey_vlbm/ani")

# Load annotations
trajs_2d    = np.load(seq_dir / "trajs_2d.npy")     # (T, N, 2)
trajs_3d    = np.load(seq_dir / "trajs_3d.npy")     # (T, N, 3)
vis         = np.load(seq_dir / "visibilities.npy")  # (T, N)
intrinsics  = np.load(seq_dir / "intrinsics.npy")    # (T, 3, 3)
extrinsics  = np.load(seq_dir / "extrinsics.npy")    # (T, 4, 4)

# Load an image and depth map
frame_idx = 0
rgb = Image.open(seq_dir / "rgbs" / f"rgb_{frame_idx:05d}.jpg")
depth_npz = np.load(seq_dir / "depths" / f"depth_{frame_idx:05d}.npz")
depth = depth_npz['depth']  # float16 array (H, W)

# Load scene info
with open(seq_dir / "scene_info.json", 'r') as f:
    scene_info = json.load(f)

print(scene_info)

Citation

Please cite the original PointOdyssey dataset when using the converted data:

@inproceedings{zheng2023pointodyssey,
  title={PointOdyssey: A Large-Scale Synthetic Dataset for Long-Term Point Tracking},
  author={Zheng, Yang and Harley, Adam W. and Shen, Bokui and Wetzstein, Gordon and Guibas, Leonidas J.},
  booktitle={ICCV},
  year={2023}
}