SpineNet is a CVPR 2020 paper for object detection. This project is a kind of implementation of SpineNet using mmdetection.

It is based on the

• the paper SpineNet: Learning Scale-Permuted Backbone for Recognition and Localization
• official TensorFlow implementation

Note: The parameters and flops are a little different from paper, so I think there are some difference between my code and official's code. More information about models can see in MODEL_DETAILS.md

1. Install mmdetection

   This implementation is based on mmdetection(v1.1.0+8732ed9). Please refer to INSTALL.md for installation and dataset preparation.

2. Copy the codes to mmdetection directory

   cp -r mmdet/ ${MMDETECTION_PATH}/
   cp -r configs/ ${MMDETECTION_PATH}/

3. Prepare data

   The directories should be arranged like this:

   >   mmdetection
   >     ├── mmdet
   >     ├── tools
   >     ├── configs
   >     ├── data
   >     │   ├── coco
   >     │   │   ├── annotations
   >     │   │   ├── train2017
   >     │   │   ├── val2017
   >     │   │   ├── test2017
   

4. Train D0 with 4 GPUs

   CONFIG_FILE=configs/spinenet/spinenet_49_B_8gpu.py
   ./ tools/dist_train.py ${CONFIG_FILE} 4

5. Calculate parameters and flops

    python tools/get_flops.py ${CONFIG_FILE} --shape $SIZE $SIZE

6. Test

   python tools/test.py ${CONFIG_FILE} ${CHECKPOINT_FILE} --out  ${OUTPUT_FILE} --eval bbox

More usages can reference mmdetection documentation.

• [2020-05-30] Add SpineNet-49S results.
• [2020-05-13] Update codes and results.
• [2020-05-02] Create this repository.