ZhengGuangze/PointOdyssey_vlbm_old

PointOdyssey to VLBM Format Conversion Report

This document summarizes the process and results of converting the PointOdyssey training dataset to the Visual Lattice Boltzmann Model (VLBM) format.

Conversion Overview

The PointOdyssey train split was converted using a multi-processed Python script (pointodyssey2vlbm.py). The conversion involved transforming source RGB images, 16-bit depth maps, and coordinate annotations into the standardized format used by the VLBM dataset loader.

Processing Steps

1. 1.RGB Images: Original JPG images were re-saved with a quality setting of 95 to ensure consistency with other VLBM datasets (like MVS-Synth).
2. 2.Depth Maps: Source 16-bit PNG depth maps (in millimeters) were converted to float32 (in meters), and stored as compressed .npz files in float16 precision to balance accuracy and storage.
3. 3.Annotations: Coordinate data was mapped from the original anno.npz to the VLBM annotations.npz format:
4. 4.trajs_3d and trajs_2d: Trajectory coordinates (float32).
5. 5.visibilities: Renamed from visibs for compatibility.
6. 6.intrinsics and extrinsics: Camera matrices (float32), tiled per-frame where necessary.
7. 7.Metadata: A scene_info.json was generated for each sequence containing frame counts, image dimensions, and depth ranges.

Results Summary

• —Total Successful Sequences: 109
• —Skipped Sequences: 22 (Incomplete or corrupted anno.npz files lacking 3D trajectory data).
• —Total Frames: 205,847
• —Target Directory: data/pointodyssey_vlbm

Data Format & Composition

Storage & Statistics

Repository Upload Guide

The processed dataset in data/pointodyssey_vlbm is ready for upload to Hugging Face datasets. It follows the same directory structure as data/kubric and data/mvs-synth_vlbm, making it directly compatible with the vlbm training pipeline.