cam_3d_reconstruction

This is a series tools and models for reconstructing dense pointcloud using camera images recorded when driving.

Module

prepare_dataset

prepare data (images, poses) for depth prediction and pointcloud alignment

• Undistort, crop images
• Make train and validation files list
• Extract relative and absolute poses from GPS information
• Remove small objects in an object instance mask
• Remove static scenes in a video

postprocess_data

• npy file (predicted mask) to tiff
• Given rgb images and absolute depth for each pixel, obatin a colorful pointcloud in camera frame
• Align isolated pointclouds with absolute poses

monodepth2_autowise

monodepth2 with:

• autowise_dataset.py
• inference_depth_pose.py
• translation_loss to fit the scale (trainer.py)
• training with mask

doc

Documenting issues about pipeline of cam_3d_reconstruction.