/RCNN

a pytorch-based reproduction of R-CNN model for Object Detection

Primary LanguagePythonMIT LicenseMIT

R-CNN PyTorch Reproduction

This is a reproduction project for 2014 CVPR paper 《Rich feature hierarchies for accurate object detection and semantic segmentation》 based on PyTorch. We reproduced R-CNN on PASCAL VOC dataset.

Dependency

To install requiring packages, use

pip install -r requirements.txt

Usage & Configuration

To train the R-CNN model, you can use scripts stored in ./script, which is made up using configuration as below. The configuration part for this repo used hydra, and the default configuration files is stored in ./config. You can simply override any configuration by running.

python train_step2.py data.train.batch_size=512

which makes your size of mini-batch in training data 512.

To check your configuration before experiments, simply run:

python train_step2.py --cfg job

And your experimental logs and configurations will be stroed in ./outputs/some-directory, you can set the specific directory by running:

python train_step2.py hydra.run.dir=./outputs/your-own-directory

Evaluation

The evaluation part of this project is from Object-Detection-Metrics projects.

You can check the evaluation results by first running ./script/test.sh, which will run './evaluation/pascalvoc.py'