/MonoJSG

Code for MonoJSG: Joint Semantic and Geometric Cost Volume for Monocular 3D Object Detection

MonoJSG

MonoJSG: Joint Semantic and Geometric Cost Volume for Monocular 3D Object Detection

Quick Start

Installation

Step 0. Follow the readme to setup the environment for MMDet3D. Step 1. Setup the KITTI dataset and preprocess the data as in MMDet3D. Step 2. Build DCNv2 by bash make.sh.

Training

Step 0. Pretrain the detection network with NOCS

bash tools/dist_train.sh configs/centernet3d/nocs/centernet3d_nocs_kitti.py 1

Step 1. Finetune the pretrained model with MonoJSG.

bash tools/dist_train.sh configs/centernet3d/two_stages/centernet3d_monojsg_kitti.py 1 --cfg-options load_from=$pretrained_model