olivermao/sceneflow dataset hosted on Hugging Face and contributed by the HF Datasets community

This dataset contains 9 x 9 Views of frame-pairs from several shots from the Blender Open Source movie, "Spring". Ground Truth scene flow and depth-maps are stored as .exr files.

The files are grouped according to scene. There are six scenes: - 01_025_A - 02_020_A - 02_040_A - 02_055_A - 05_030_A - 08_040_A

A brief summary of important meta data is included below:

Scene Frame Pair Spacing (mm) Focal length (mm) Frame Rate (fps) 01_025_A 840, 841 0.5 35 96 02_020_A 510, 511 3.0 21 72 02_040_A 360, 361 9.0 16 72 02_055_A 2064, 2065 1.0 65 384 05_030_A 6390, 6391 3.0 21 360 08_040_A 145, 146 110.0 18 24

The original framerate was 24 fps.

A file containing meta data for each scene called meta_data.json is located within the relevant .zip file for each scene. It contains: - Focal Length (mm) - Sensor Width (px) - Sensor height (px) - Horizontal Resolution (px) - Vertical resolution (px) - Number of views in the x direction - Number of views in the y direction - Baseline or spacing between cameras (mm)

The coordinate for each view is given by the folder name Cam (x, y). Ground truth data is given for the central view in the folder Cam (0, 0). Each image is saved as a .png file.

The ground truth flow fields in the .exr files are stored as a 3D array with dimensions (height, width, 3). At a given row and column, the motion vector is given in the order (z, y, x). For example, [84, 42, 0] has the component of the flow field in the z direction at row 84, column 42 and, [42, 84, 2] has the x component of the flow field at row 42, column 84. Each component of the vector field is given in the unit, metres per frame (m / frame).

If you use this work please cite:

@phdthesis{Gray_2024, author = {James L. Gray}, title = {Gradient Consistency: A New Take on Variational Optical Flow and Disparity Estimation}, school = {University of New South Wales}, address = {Sydney, NSW}, year = {2024}, month = {jul} }

Self-driving dataset for scene flow estimation

This torrent contains the "Optical Flow" data for a one-sixteenth-baseline version of the "FlyingThings3D" dataset from the CVPR 2016 paper "A Large Dataset to Train Convolutional Networks for Disparity, Optical Flow, and Scene Flow Estimation" by Mayer et al. ().

The Spring dataset contains files for scene flow, optical flow and stereo estimation. For easier handling, we organized them into sub-directories: train split: train_frame_left.zip: left camera frames train_frame_right.zip: right camera frames train_disp1_left.zip: left-to-right disparity in the reference frame train_disp1_right.zip: right-to-left disparity in the reference frame train_disp2_FW_left.zip: left-to-right disparity in the future/forward target frame train_disp2_BW_left.zip: left-to-right disparity in the past/backward target frame train_disp2_FW_right.zip: right-to-left disparity in the future/forward target frame train_disp2_BW_right.zip: right-to-left disparity in the past/backward target frame train_flow_FW_left.zip: left forward optical flow train_flow_BW_left.zip: left backward optical flow train_flow_FW_right.zip: right forward optical flow train_flow_BW_right.zip: right backward optical flow train_cam_data.zip: camera data: intrinsics, extrinsics, focal distance train_maps.zip: additional maps: detail, match, rigid, sky test split: test_frame_left.zip: left camera frames test_frame_right.zip: right camera frames test_cam_data.zip: camera data: intrinsics File formats: images and maps are given in png format optical flow files are given in HDF5 file format and named .flo5 disparity files are given in HDF5 file format and named .dsp5

Graph matching problems for large displacement optical flow of RGB-D images.

Results of EPE by adding different number of PA modules on Scene flow dataset.

The RobustSpring dataset contains the image corruption data files for scene flow, optical flow and stereo estimation with the Spring dataset. Note that this repository contains only the Spring test data files. For easier handling, we organized them into sub-directories by image corruption type: brightness.zip : brightness image corruption contrast.zip : contrast image corruption defocus_blur.zip : defocus_blur image corruption elastic_transform.zip : elastic_transform image corruption fog.zip : fog image corruption frost.zip : frost image corruption gaussian_blur.zip : gaussian_blur image corruption gaussian_noise.zip : gaussian_noise image corruption glass_blur.zip : glass_blur image corruption impulse_noise.zip : impulse_noise image corruption jpeg_compression.zip : jpeg_compression image corruption motion_blur.zip : motion_blur image corruption pixelate.zip : pixelate image corruption rain.zip : rain image corruption saturate.zip : saturate image corruption shot_noise.zip : shot_noise image corruption snow.zip : snow image corruption spatter.zip : spatter image corruption speckle_noise.zip : speckle_noise image corruption zoom_blur.zip : zoom_blur image corruption Each image corruption folder is internally organized as follows: test : Indicates that this is the test proportion of the Spring dataset

Scene flow estimation on point clouds guided by optimal transport.

Analysis of performance with different numbers of hourglasses on Scene Flow [19].

This torrent contains the "Disparity Change" data for a one-sixteenth-baseline version of the "FlyingThings3D" dataset from the CVPR 2016 paper "A Large Dataset to Train Convolutional Networks for Disparity, Optical Flow, and Scene Flow Estimation" by Mayer et al. ().

FlyingThings3D is a synthetic dataset for optical flow, disparity and scene flow estimation. It consists of everyday objects flying along randomized 3D trajectories. We generated about 25,000 stereo frames with ground truth data. Instead of focusing on a particular task (like KITTI) or enforcing strict naturalism (like Sintel), we rely on randomness and a large pool of rendering assets to generate orders of magnitude more data than any existing option, without running a risk of repetition or saturation.

Learning scene flow in 3D point clouds.

A neural scene flow fields for space-time view synthesis of dynamic scenes.

The results of PSM-Net [21] and GWC-Net [13] are trained with published code with our batch size, evaluation settings for fair comparison.

A Dynamic-weather Driving Dataset designed specifically to enhance autonomous driving systems' adaptability to diverse weather conditions. Includes stereo image pairs (left and right RGB) with ground truth depth, optical flow and delta disparity.

Quantitative evaluation of Fast-GFM on ETH3D [33] and Middlebury [34].

This torrent contains the "Cleanpass" image data for a one-sixteenth-baseline version of the "FlyingThings3D" dataset from the CVPR 2016 paper "A Large Dataset to Train Convolutional Networks for Disparity, Optical Flow, and Scene Flow Estimation" by Mayer et al. ().

The results of the ablation comparison on Scene Flow.

The experiment of ASPP structure on Scene Flow.