fulkast/drs-vpt-demo-data

DRS-VPT demo data

One preprocessed sample per dataset (plus an iPhone LiDAR capture of our own, as a single frame and as a 37-frame stream), used by the DRS-VPT demo (scripts/demo.py) and its regression tests, and a regression set of 10 KITTI and 10 nuScenes benchmark samples (regression/). The code downloads the samples it needs automatically; nothing has to be fetched by hand.

Non-commercial use only. Each subset keeps the licence of its source dataset (see the table and LICENSE.md). The 7scenes subset is under the Microsoft Research License Agreement, not Creative Commons; see 7scenes/NOTICE.md. If you use a sample, cite the original dataset.

Layout

<dataset>/
  image.{png,jpg}  query image, unmodified from the source dataset
  K.npy            3x3 pinhole intrinsics of image (float32)
  cloud.npy        N x 3 map point cloud (float32, metres), exactly as fed to the model
  meta.json        grid_size (voxel size for the cloud encoder), provenance, ground truth

iphone_stream holds many query frames against one map (scripts/demo.py --sample iphone_stream):

iphone_stream/
  cloud.npy        N x 3 map point cloud (float32, metres)
  frames/*.jpg     query images
  meta.json        grid_size, provenance, and "frames": [{"image", "K", "T_c2w"}]

Subset	Query	Map cloud	Points	Ground truth in meta.json	Source licence
kitti	VP2P eval split, sample 539	LiDAR scan, after the eval protocol's seeded random rigid transform	123k	P as returned by the eval loader	CC BY-NC-SA 3.0 (modified cloud: CC BY-NC-SA 4.0)
7scenes	stairs / seq-01 / frame-000000	TSDF points fused from the train split, centroid-subtracted	722k	T_c2w (camera-to-world, in the cloud frame)	MSR-LA, non-commercial
12scenes	apt2/living, first test frame	mliving.ply vertices, centroid-subtracted	3.3M	none (tracking failed for this frame in the source data)	CC BY-NC-SA 4.0
iphone	iPhone 14 Pro, Polycam keyframe 18 (portrait frames stored sideways, as captured)	High-confidence LiDAR depth of 32 other keyframes (query and +/-2 neighbours excluded), centroid-subtracted	1.16M	T_c2w (ARKit pose, converted to OpenCV axes)	CC BY-NC-SA 4.0
iphone_stream	iPhone 14 Pro, all 37 Polycam keyframes	High-confidence LiDAR depth of all 37 keyframes (includes each query's own depth, so GT errors are optimistic), centroid-subtracted	1.36M	T_c2w per frame (ARKit pose, OpenCV axes)	CC BY-NC-SA 4.0

centroid in meta.json is the offset subtracted from the original world frame.

Regression set

regression/{kitti,nuscenes}/<name>/ holds the benchmark samples the DRS-VPT regression tests localize (tests/test_regression.py), in the single-frame layout above. They were picked by scripts/export_regression_set.py across sequences and the error distribution of a full benchmark run, all localized within the recall thresholds. meta.json also records source_image (relative to the benchmark data root), the pick's group and reason, and the benchmark errors of that run; T_c2w is the ground truth.

Subset	Samples	Map cloud	Source licence
regression/kitti	10 from the VP2P eval split (sequences 09 and 10)	LiDAR scan in the camera frame, after the eval protocol's seeded random rigid transform	CC BY-NC-SA 3.0 (modified clouds: CC BY-NC-SA 4.0)
regression/nuscenes	10 from the nuScenes v1.0 test split, as built by data_processing/nuscenes_preprocess.py	LiDAR scan in the camera frame, after the eval protocol's seeded random rigid transform	CC BY-NC-SA 4.0

Sources

• KITTI: Geiger et al., Vision meets Robotics: The KITTI Dataset, IJRR 2013.
• 7-Scenes: Shotton et al., Scene Coordinate Regression Forests for Camera Relocalization in RGB-D Images, CVPR 2013.
• 12-Scenes: Valentin et al., Learning to Navigate the Energy Landscape, 3DV 2016.
• nuScenes: Caesar et al., nuScenes: A Multimodal Dataset for Autonomous Driving, CVPR 2020.