datasets
Training and evaluation data, with the modality, task and licence stated up front. Listed live from the Hugging Face Hub.
MARS_benchmark
MARS Benchmark Dataset
A curated benchmark dataset for evaluating forward-warped novel view synthesis in autonomous driving scenarios.
30 scene pairs extracted from the MARS dataset (Ann Arbor, Michigan)
Depth estimated with Depth Anything V3 (DA3-GIANT)
Forward warping performed with the cosmos_predict2 warping module
Folder Structure
<pair_name>/
images/ # Source images (ref + tgt_subset), PNG
depth/ # DA3 depth maps for ref +… See the full description on the dataset page: https://huggingface.co/datasets/enrue1893/MARS_benchmark.MARS_BENCHMARK_RAW
MARS short_seq — Folder Structure
Each short_sequence_xxxx/ is organized into three folders:
Folder
Meaning
Contents
gt/
Ground Truth
frames/ (real rendered frames), dynamic_masks/ (dynamic-object masks)
ref/
Reference (input / conditioning views)
images/, depth/, intrinsics/, world_extrinsics/ — ref_* views only
gen_track/
Generation target (what we aim to generate)
warped_frames/, warped_video.mp4, extrinsics/, intrinsics/, warp_mask/
Other per-sequence… See the full description on the dataset page: https://huggingface.co/datasets/enrue1893/MARS_BENCHMARK_RAW.MARS-Bench