Spatial-AI

About

To improve depth estimation and disparity map generation using OAK-D Pro camera. To justify a method to get highly accurate depth map by utilizing traditional knowledge and modern deep learning techniques.

Methodolgy

Part 1 : Implementing Pre + Post Processing

Performing pre-processing on stereo left and stereo right image and then implementing stereo rectification as well as triangulation method increases depth perception of camera manifolds. As well as reduce its noise. Preprocessing on images with certain touch on it with inbuilt OAK D post processing filters has improved its depth a lot.

Part 2 : Applying Midas Model On RGB Video

MiDAS is a pretrained model which improves Monocular Depth Estimation (MDE) of monocular RGB video. It generates a state of the art image. The success of monocular depth estimation relies on large and diverse training sets

Part 3 : Fusing Stereo Disparity Map and MDE Disparity Map

In this approach we fuse disparity map generated by OAK-D using stereo cameras and disparity map generated using MiDAS model (MDE) on rgb video. This method aims to combine excellent features of Stereo as well as Monocular Depth Estimations and reduce noise generated by one disparity map by superimposing quality of other disparity map.

Results

Github Link

https://github.com/Smit1603/Spatial-AI