Backbone | AP | [email protected] | inference time |
---|---|---|---|
Mobilenet v1 | 31.7 | 49.8 | 36 ms |
Shufflenet v2 | 29.3 | 46.7 | 31 ms |
- Inference time was measured on images of size 896x640 and using NVIDIA GTX 1080 Ti.
- Average precision was computed on COCO val 2017 dataset.
- During the evaluation I resize smallest dimension to be equal to 640 (while keeping the aspect ratio).
- You can get the pretrained models from here.
You can test them usinginference/just_try_detector.ipynb
.
- Prepare dataset. See
data/README.md
. - Edit a json configuration file and then just run
python train.py
. - Then use
create_pb.py
to get a frozen inference graph. - Use notebooks in
inference
folder to test the trained detector.
- tensorflow 1.12
- tqdm, Pillow, numpy, matplotlib
This implementation is based on Tensorflow Object Detection API.