EfficientObjectDetection
Our code and experiments around the paper https://arxiv.org/abs/1912.03966.
We consider a simplified version of the framework presented in the paper. In particular, we train only one policy network which distinguish between images which need to be elaborated with a large and accurate detector and images which is sufficient to be processed with a small and coarse detector. In our experiments we deal with COCO image detection dataset and don't split images into patches. Our overall inference procedure is presented in the image
Requirements
-
mmdetectionframework. For the installation instruction see the original documentation -
COCO dataset. The dataset details and distributed files could be found on the official website
To perform the inference one need to use the coarse and fine level detectors form mmdetection. The links to the corresponding configs as well as checkpoints are available in the table below
| network | link to the config | link to checkpoint |
|---|---|---|
| RegNetX-4GF-FPN | fine level config | fine model checkpoint |
| RegNetX-400MF-FPN | coarse level config | coarse model checkpoint |
Download link for the ResNet18 policy network checkpoint.
Inference
The demo of the inference is available in ./inference_demo.ipynb notebook. One need to set up all necessary pathes to datasets, checkpoints and configs and substitute demo agent with trained ResNet18 policy network.
Credits
-
Original code by the authors of https://arxiv.org/abs/1912.03966
-
Celebrated mmdetection framework