(Note: Upper part is Disparity Map and bottom part is Object detection + Depth Estimation(z=?))

Please check my Medium Blog for more information

Full Video output is shared at Link

Incase of Stereo Setup, Depth estimation is dependent on disparity map.

[PointCloud Output]

• Download Pre-Trained model which i shared at Download Link

  Place it inside root folder and update the path in the config.py.

  RAFT_STEREO_MODEL_PATH = "pretrained_models/raft_stereo/raft-stereo_20000.pth"
  FASTACV_MODEL_PATH = "pretrained_models/fast_acvnet/kitti_2015.ckpt"
  ...
  

• Download Yolo for object detection.I shared it at Download Link.

Download Kitti Dataset from Download Link

• Download Left/Right Images: Download stereo 2015/flow 2015/scene flow 2015 data set (2 GB)
• Download Calibration files: Download calibration files (1 MB)

Keep these files in some path and update config.py

[config.py]
KITTI_CALIB_FILES_PATH=".../kitti_stereo_2015/data_scene_flow_calib/testing/calib_cam_to_cam/*.txt"
KITTI_LEFT_IMAGES_PATH=".../kitti_stereo_2015/testing/image_2/*.png"
KITTI_RIGHT_IMAGES_PATH=".../kitti_stereo_2015/testing/image_3/*.png"
...

Run "python3 demo.py" change the configuration in config.py in order to run different architecture such as BGNet, CreStereo, RAFT-Stereo etc.

KITTI_CALIB_FILES_PATH=".../kitti_stereo_2015/data_scene_flow_calib/testing/calib_cam_to_cam/*.txt"
KITTI_LEFT_IMAGES_PATH=".../kitti_stereo_2015/testing/image_2/*.png"
KITTI_RIGHT_IMAGES_PATH=".../kitti_stereo_2015/testing/image_3/*.png"

RAFT_STEREO_MODEL_PATH = "pretrained_models/raft_stereo/raft-stereo_20000.pth"
FASTACV_MODEL_PATH = "pretrained_models/fast_acvnet/kitti_2015.ckpt"
DEVICE = "cuda"

# raft-stereo=0, fastacv-plus=1, bgnet=2, gwcnet=3, pasmnet=4, crestereo=5, hitnet=6, psmnet=7
ARCHITECTURE_LIST = ["raft-stereo", "fastacv-plus", "bgnet", 'gwcnet', 'pasmnet', 'crestereo', 'hitnet', 'psmnet']
ARCHITECTURE = ARCHITECTURE_LIST[1]
SAVE_POINT_CLOUD = 0
SHOW_DISPARITY_OUTPUT = 1
SHOW_3D_PROJECTION = 0

Different state of the art (SOTA) deep learning based architetures are proposed to solve disparity and are given below:

Here is the profiling data:

Here is the inference time on Nvidia-2080Ti

• Issue with HitNet Implementation.

Thanks to the authors of fastacv-plus, bgnet, gwcnet, pasmnet, crestereo, hitnet, psmnet and raft-stereo for their opensource code.

• https://github.com/princeton-vl/RAFT-Stereo.git.
• https://github.com/gangweiX/Fast-ACVNet.
• https://github.com/3DCVdeveloper/BGNet.
• https://github.com/megvii-research/CREStereo.
• https://github.com/ibaiGorordo/HITNET-Stereo-Depth-estimation.
• https://github.com/xy-guo/GwcNet.
• https://github.com/JiaRenChang/PSMNet.
• https://github.com/The-Learning-And-Vision-Atelier-LAVA/PAM/tree/master/PASMnet.

Reach me @

LinkedIn GitHub Medium