Triplet Center Loss (TCL) Solution for SHREC2018 IBR & SBR

This repo holds the code for our method Triple Center Loss(TCL) at the SHREC 2018 challenges:

• Sketch-Based 3D Scene Retrieval(SBR)
• Image-Based 3D Scene Retrieval(IBR).

The TCL method is based on the CVPR 2018 work

Xinwei He, Yang Zhou, Zhichao Zhou, Song Bai, Xiang Bai. Triplet-Center Loss for Multi-View 3D Object Retrieval. CVPR 2018. [pdf]

For a detailed description of our submitted method, you can also refer to the technical report and find our method scription in Sec 4.2.

SHREC’18 Track: 2D Image-Based 3D Scene Retrieval. [pdf]

SHREC’18 Track: 2D Scene Sketch-Based 3D Scene Retrieval. [pdf]

Prerequisites

Our code has been tested with Python2 + PyTorch 0.3. It should work with higher versions after minor modifications.

Usage

The code works for both the IBR and the SBR tasks.

For the IBR task, run the following command

bash run_image_based.sh $gpu $backbone $action $suffix $output_dir

Arguments:

• gpu: the GPU ID
• backbone: the backbone architecture to use. e.g. vgg11_bn, resnet50.
• action: 1 or 2. choose 1 for training and 2 for inference.
• output_dir: where to save the outputs

For the SBR task, use the script run_sketch_based.sh instead. The arguments are the same.

Related Projects

• cvpr_2018_TCL.pytorch