FasterRCNN

Faster-RCNN on COCO

This example aims to provide a minimal (<1000 lines) multi-GPU implementation of ResNet50-Faster-RCNN on COCO.

Dependencies

• TensorFlow >= 1.4.0rc0
• Install pycocotools, OpenCV.
• Pre-trained ResNet50 model from tensorpack model zoo.
• COCO data. It assumes the following directory structure:

DIR/
  annotations/
    instances_train2014.json
    instances_val2014.json
    instances_minival2014.json
    instances_valminusminival2014.json
  train2014/
    COCO_train2014_*.jpg
  val2014/
    COCO_val2014_*.jpg

minival and valminusminival are optional. You can download them here.

Usage

Change BASEDIR in config.py to /path/to/DIR as described above.

To train:

./train.py --load /path/to/ImageNet-ResNet50.npz

The code is only for training with 1, 2, 4 or 8 GPUs. Otherwise, you probably need different hyperparameters for the same performance.

To predict on an image (and show output in a window):

./train.py --predict input.jpg

Results

Mean Average Precision @IoU=0.50:0.95:

• trainval35k/minival, FASTRCNN_BATCH=256: 33.4. Takes 49h on 8 TitanX.
• trainval35k/minival, FASTRCNN_BATCH=64: 32.2. Takes 31h on 8 TitanX.

The hyperparameters are not carefully tuned. You can probably get better performance by e.g. training longer.

Notes

See Notes on This Implementation