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.
To install requiring packages, use
pip install -r requirements.txt
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
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'