anujayavidmal2002/FYP-UOM-PreviousData

FYP-UOM 4D Gaussian Splatting Training Data

Intermediate and final outputs of a Final Year Project (University of Moratuwa) that reconstructs monocular videos as 4D Gaussian Splatting (4DGS) scenes and splits each reconstruction into static and dynamic Gaussians using motion and object priors.

There are 64 videos from public video-segmentation benchmarks. For each one, the repo holds the sampled frames, the estimated geometry and cameras, optical flow, SAM2-based motion and object priors, the trained 4DGS checkpoints and the final static/dynamic split. Total size is about 131 GB in about 43k files.

Source videos

Folder	Source	Scenes
scenes/davis_2017_full_raw/	DAVIS 2017 (train/val)	21: bear, blackswan, boat, car-shadow, disc-jockey, dog, flamingo, kite-surf, kite-walk, libby, miami-surf, motocross-bumps, night-race, paragliding-launch, paragliding, planes-water, rollerblade, soccerball, surf, swing, tennis
scenes/davis_2017_test_dev_raw/	DAVIS 2017 test-dev	10: deer, giant-slalom, helicopter, lock, man-bike, orchid, planes-crossing, skate-jump, slackline, tennis-vest
scenes/davis_2017_test_challenge_raw/	DAVIS 2017 test-challenge	4: dog-control, running, swing-boy, turtle
scenes/fbms_59_raw/	FBMS-59	29: camel01, cats01–07, dogs01–02, horses02–06, lion01–02, marple*, people1, rabbits01–05, tennis

The raw frames were re-encoded as 24 fps videos (<scene>_raw_24fps.mp4), and 32 frames were sampled uniformly across each whole video. planes-crossing has only 31 unique frames, so all 31 are used.

Per-scene layout

Each scene is in scenes/<dataset>/<scene>_raw_24fps.mp4/:

Path	Contents
frames.json, source_frames/	Sampled frame indices/timestamps and the source frames
fourrc.npz	Native 4RC predictions (point maps, poses, intrinsics, confidence, trajectories)
rgb/, depth/, confidence/, pointmaps/	Aligned per-frame RGB, depth, confidence and world-point maps
cameras.npz, initial_points.ply	Camera arrays and the initial point cloud for 4DGS
flow/	Bidirectional RAFT optical flow
priors/	Camera-corrected residual motion and per-frame SAM2 masks
object_priors/	SAM2 object discovery/tracking, final dynamic targets and fallback masks
4dgs/	4DGS reconstruction (3k coarse + 14k fine iterations): point clouds and deformation networks
4dgs/dynamic_split_corrected/	Final static/dynamic PLYs, membership probability maps, hard masks, videos and prior-agreement reports (threshold 0.98)
logs/	Per-step logs

Files are mostly .png, .npy, .npz, .ply, .pth, .json, .csv and .mp4.

Pipeline code

The root of this repo has the batch runner (run_all_videos.sh, pipeline_helpers.py, pipeline_state.py), the list of input videos (videos.txt), the run settings and the pipeline logs. PIPELINE_README.md describes the execution order and every output in detail.

Download

# Everything (~131 GB)
hf download anujayavidmal2002/FYP-UOM-PreviousData --repo-type dataset --local-dir ./data

# A single scene
hf download anujayavidmal2002/FYP-UOM-PreviousData --repo-type dataset \
  --include "scenes/davis_2017_full_raw/bear_raw_24fps.mp4/*" --local-dir ./data

License and terms

This repo contains frames derived from DAVIS 2017 and FBMS-59. Those frames stay under the terms of their original datasets, so read LICENSE_NOTICE.md and cite the original datasets if you use this data. The derived outputs (geometry, priors, reconstructions) are shared for non-commercial research use.

Citation

If you use this data, please cite the source datasets:

@article{pont-tuset2017davis,
  title   = {The 2017 DAVIS Challenge on Video Object Segmentation},
  author  = {Pont-Tuset, Jordi and Perazzi, Federico and Caelles, Sergi and Arbel{\'a}ez, Pablo and Sorkine-Hornung, Alexander and Van Gool, Luc},
  journal = {arXiv:1704.00675},
  year    = {2017}
}

@article{ochs2014fbms,
  title   = {Segmentation of Moving Objects by Long Term Video Analysis},
  author  = {Ochs, Peter and Malik, Jitendra and Brox, Thomas},
  journal = {IEEE Transactions on Pattern Analysis and Machine Intelligence},
  volume  = {36},
  number  = {6},
  pages   = {1187--1200},
  year    = {2014}
}