Stereo Correspondence and Reconstruction of Endoscopic Data Challenge

This is a project about the "Stereo Correspondence and Reconstruction of Endoscopic Data Challenge", which includes data download and processing, the implementation of multiple stereo matching models, and the depth estimation(reconstruction) results display.

How to use

Environment

• python 3.9
• Pytorch '1.10.2+cu111'

Data Preparation

Download

Download SCARED Dataset from SCARED Datasets

Preprocessing

Run the script ./scripts/preprocessing.sh to unzip, extract images from video, rectify the images and ground truth.Note that you can rectify the images only, but need reverse the rectify from the predicted(rectified) to original(unrectified).

Run the script ./scripts/get_csv.sh to get a .csv file which contains and organizes the path information of the necessary data.By default the .csv files will saved to ./csvfiles.

Training

Here are five available models,you can choose one of them['stackhourglass', 'basic', 'constancy', 'gwc_g', 'gwc_gc'] by set the --model argument in the ./scripts/train.sh.Remember update other command arguments.

Run the script ./scripts/train.sh to train the model you choose on the SCARED dataset.

Evaluation

Run the script ./scripts/test.sh to evaluate the predictions, remember specify a savedir to store the results.

Representation

The details of basic and stackhourglassmodel can be found in "Pyramid Stereo Matching Network".

The constancy means using feature constancy as a loss term,and the gwc_g and gwc_gcmeans using group-wise correlation to build the cost volume.

The depth estimation of the stackhourglass, constancy, gwc_gc and the reconstructed surface are shows as below:

​