This project is intended to be a TensorFlow implementation of the RefineDet object detection architecture, from the paper Single-Shot Refinement Neural Network for Object Detection (Zhang et al.), originally made in Caffe.
I mostly worked on this project during spare time, and as a way to learn more about TensorFlow in practice. Slight details might deviate from the official Caffee implementation, but all the main ideas are present.
The code has been completely adapted to work with TensorFlow 2.0, and it only uses TF operations, instead of python logic + tf.py_function as was the case in the initial version of this project, which gave a significant boost in training speed.
I am releasing an inference demo script, as well as a weights file, trained on the 20 class PASCAL VOC 07+12 trainval datasets (Everingham et al.), with 320 x 320 input size. These weights are achieving around 80% mean AP on the VOC 2007 test set, which is about the same as reported on the paper. But note that I used my own custom Python implementation of the VOC Mean Average Precision, which might deviate slightly from the official Matlab version.
In order to run the demo script on the sample images, first download the pretrained weights file from here into ./weights, and then simply run:
python demo.pyand it will perform inference on a few Pascal VOC test images inside the samples directory. The results will be displayed on the screen and saved inside samples/detections.
If you want to perform evaluation on the whole Pascal VOC 2007 test set, as done in the paper, first download the 2007 test partition of the dataset, and extract it inside ./data . You can find the the main Pascal VOC files in the pjreddie mirror.
Assuming you already have the weights file inside ./weights, you just have to run.
python eval_refinedet_voc.pyTake a look in this file to see optional parameters.
If you want to train from scratch, you will need to download the pretrained weigths for the VGG16 backbone CNN from here. These weights were provided by the popular pierluigiferrari's ssd_keras repository.
You will also need the Pascal VOC 2007 and 2012 trainval partitions, which you can also find in the pjreddie mirror.
If you have the pretrained weights inside ./weights and the VOC data extracted inside ./data, then training is just running the train script:
python train_refinedet_voc.pyTake a look in this file to see optional parameters.
If you want to use this implementation in another project, simply import the models.RefineDetVGG16 class. Take a look at the file for clarification on the available parameters.
If you want to implement a different RefineDet based detector, you can extend models.RefineDetBase class and implement your own forward pass logic inside the call method. Follow the example in the RefineDetVGG16 class.
If you want to retrain it on another dataset, I believe the files are easy enough to understant what is happening, so use the files train_refinedet_voc.py and eval_refinedet_voc.py as references for training and evaluation, respectively. If you are going to use the VGG16 CNN backbone (Simonyan et al.), download the pretrained weights from the
Here are some sample detections on the PASCAL VOC 2007 test set.
- Original RefineDet Paper: Zhang, Shifeng and Wen, Longyin and Bian, Xiao and Lei, Zhen and Li, Stan Z., Single-Shot Refinement Neural Network for Object Detection, CVPR 2018
- Official Caffe implementation: https://github.com/sfzhang15/RefineDet
- I also took some inspiration from some RefineDet implementations other than the official one, specially the one from luuuyi. If you prefer Pytorch, check out his implementation.