FasterRCNN

Faster-RCNN / Mask-RCNN on COCO

This example aims to provide a minimal (1.3k lines) multi-GPU implementation of Faster-RCNN / Mask-RCNN (without FPN) on COCO.

Dependencies

• Python 3; TensorFlow >= 1.4.0
• pycocotools, OpenCV.
• Pre-trained ResNet 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 config in config.py:

1. Set BASEDIR to /path/to/DIR as described above.
2. Set MODE_MASK to switch Faster-RCNN or Mask-RCNN.

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.

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

./train.py --predict input.jpg --load /path/to/model

Evaluate the performance of a model and save to json. (Pretrained models can be downloaded in model zoo:

./train.py --evaluate output.json --load /path/to/model

Results

Models are trained on trainval35k and evaluated on minival using mAP@IoU=0.50:0.95. MaskRCNN results contain both bbox and segm mAP.

Backbone	FASTRCNN_BATCH	resolution	mAP (bbox/segm)	Time
Res50	64	(600, 1024)	33.0	22h on 8 P100
Res50	256	(600, 1024)	34.4	49h on 8 M40
Res50	512	(800, 1333)	35.6	55h on 8 P100
Res50	256	(800, 1333)	36.9/32.3	39h on 8 P100
Res101	512	(800, 1333)	40.1/34.4	70h on 8 P100

Note that these models are trained with a larger ROI batch size than the paper, and get about 1mAP better performance.

Notes

See Notes on This Implementation