Implementation of the paper “CNN-BASED OBJECT DETECTION AND DISTANCE PREDICTION FOR AUTONOMOUS DRIVING USING STEREO IMAGES”.
imagenet pretrain (cls100)
sceneflow pretrain
kitti 2d stereo
It describes how to pretrain a stereoneck and train the entire network end-to-end.
It will run without problems in the latest version of PyTorch.
The version at the time we tested is 1.10.0.
Modify the paths inside the CFG file appropriately.
The 2D bounding box and distance labels are enclosed in the gendata folder. Unzip it.
Code to convert the 3D bounding box of the KITTI form to a 2D bounding box and distance is included in result_fromkitti.py.
The code to output the 2D bounding box and distance predicted by the network to a txt file is included in trainkitti2d_.py.
Result using pseudo LiDAR points generated from disp map of PSMNet: CIA-SSD was used for 3Ddet network.
CIA-SSD
PSMNet
YOLOStereo3D: A Step Back to 2D for Efficient Stereo 3D Detection
Dist-YOLO: Fast Object Detection with Distance Estimation
Improved Stereo Matching Accuracy Based on Selective Backpropagation and Extended Cost Volume