CenterNet TensorFlow 1.14

This repository contains the implementation of Centernet in TensorFlow 1.14

• CenterNet: Keypoint Triplets for Object Detection, Kaiwen Duan et al. 2019

Preparation

Please follow the instructions in the official repo to install: NMS, MS COCO APIs and MS COCO Data. There is no need to compile corner pooling layers, as this repo contains the code for all the pooling layers directly implemented in TensorFlow 1.14.

Docker

Build the docker image using following command:

docker build --tag "cn:tf1.14" --network=host .

Once the docker image is build, use following command to run the training:

docker run -it --rm --runtime=nvidia --name <your_name>_train --network host -v /data:/data -v <your_workspace_path>/centernet:/code cn:tf1.14 python -u /code/main.py --cfg_file /code/config/config_docker.yaml

All the training metrics can be visualized on the tensorboard with the following command:

docker run -it --rm --runtime=nvidia --name <your_name>_log --network host -p <your_port>:<your_port> -v <your_workspace_path>/centernet:/code cn:tf1.14 tensorboard --port <your_port> --logdir /code/output/logs