FasterRCNN

Faster R-CNN / Mask R-CNN on COCO

This example provides a minimal (2k lines) and faithful implementation of the following object detection / instance segmentation papers:

• Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks
• Feature Pyramid Networks for Object Detection
• Mask R-CNN
• Cascade R-CNN: Delving into High Quality Object Detection

with the support of:

• Multi-GPU / multi-node distributed training, multi-GPU evaluation
• Cross-GPU BatchNorm (aka Sync-BN, from MegDet: A Large Mini-Batch Object Detector)
• Group Normalization
• Training from scratch (from Rethinking ImageNet Pre-training)

This is likely the best-performing open source TensorFlow reimplementation of the above papers.

Dependencies

• OpenCV, TensorFlow ≥ 1.6
• pycocotools/scipy: for i in cython 'git+https://github.com/cocodataset/cocoapi.git#subdirectory=PythonAPI' scipy; do pip install $i; done
• Pre-trained ImageNet ResNet model from tensorpack model zoo
• COCO data. It needs to have the following directory structure:

COCO/DIR/
  annotations/
    instances_train201?.json
    instances_val201?.json
  train201?/
    # image files that are mentioned in the corresponding json
  val201?/
    # image files that are mentioned in corresponding json

You can use either the 2014 version or the 2017 version of the dataset. To use the common "trainval35k + minival" split for the 2014 dataset, just download the annotation files instances_minival2014.json, instances_valminusminival2014.json from here to annotations/ as well.

Usage

It is recommended to get familiar the relevant papers listed above before using this code. Otherwise you may end up doing something unreasonable.

Train:

To train on a single machine (with 1 or more GPUs):

./train.py --config \
    BACKBONE.WEIGHTS=/path/to/ImageNet-R50-AlignPadding.npz \
    DATA.BASEDIR=/path/to/COCO/DIR \
    [OTHER-ARCHITECTURE-SETTINGS]

Alternatively, use TRAINER=horovod which supports distributed training as well, but less straightforward to run. Refer to HorovodTrainer docs for details.

All options can be changed by either the command line or the config.py file (recommended). Some reasonable configurations are listed in the table below. See config.py for details about how to correctly set BACKBONE.WEIGHTS and other configs.

Inference:

To predict on given images (needs DISPLAY to show the outputs):

./predict.py --predict input1.jpg input2.jpg --load /path/to/Trained-Model-Checkpoint --config SAME-AS-TRAINING

To evaluate the performance of a model on COCO:

./predict.py --evaluate output.json --load /path/to/Trained-Model-Checkpoint \
    --config SAME-AS-TRAINING

Several trained models can be downloaded in the table below. Evaluation and prediction have to be run with the corresponding configs used in training.

Results

These models are trained on train2017 and evaluated on val2017 using mAP@IoU=0.50:0.95. Unless otherwise noted, all models are fine-tuned from ImageNet pre-trained R50/R101 models in tensorpack model zoo, using 8 NVIDIA V100s.

Performance in Detectron can be reproduced.

Backbone	mAP (box;mask)	Detectron mAP 1 (box;mask)	Time (on 8 V100s)	Configurations (click to expand)
R50-FPN	34.8		6.5h	super quick MODE_MASK=False FRCNN.BATCH_PER_IM=64 PREPROC.TRAIN_SHORT_EDGE_SIZE=[500,800] PREPROC.MAX_SIZE=1024
R50-C4	35.6	34.8	22.5h	standard MODE_MASK=False MODE_FPN=False
R50-FPN	37.5	36.7	10.5h	standard MODE_MASK=False
R50-C4	36.2;31.8 ⬇️	35.8;31.4	23h	standard MODE_FPN=False
R50-FPN	38.2;34.8	37.7;33.9	12.5h	standard this is the default
R50-FPN	38.9;35.4 ⬇️	38.6;34.5	24h	2x TRAIN.LR_SCHEDULE=2x
R50-FPN-GN	40.4;36.3 ⬇️	40.3;35.7	29h	2x+GN FPN.NORM=GN BACKBONE.NORM=GN FPN.FRCNN_HEAD_FUNC=fastrcnn_4conv1fc_gn_head FPN.MRCNN_HEAD_FUNC=maskrcnn_up4conv_gn_head TRAIN.LR_SCHEDULE=2x
R50-FPN	41.7;36.2 ⬇️		16h	+Cascade FPN.CASCADE=True
R50-FPN-GN	46.1;40.1 ⬇️		36h (on 16 V100s)	4x+GN+Cascade+TrainAug FPN.CASCADE=True FPN.NORM=GN BACKBONE.NORM=GN FPN.FRCNN_HEAD_FUNC=fastrcnn_4conv1fc_gn_head FPN.MRCNN_HEAD_FUNC=maskrcnn_up4conv_gn_head PREPROC.TRAIN_SHORT_EDGE_SIZE=[640,800] TRAIN.LR_SCHEDULE=4x
R101-C4	40.1;34.6 ⬇️		27h	standard MODE_FPN=False BACKBONE.RESNET_NUM_BLOCKS=[3,4,23,3]
R101-FPN	40.7;36.8 ⬇️ 2	40.0;35.9	17h	standard BACKBONE.RESNET_NUM_BLOCKS=[3,4,23,3]
R101-FPN	46.6;40.3 ⬇️		64h	3x+Cascade+TrainAug FPN.CASCADE=True BACKBONE.RESNET_NUM_BLOCKS=[3,4,23,3] TEST.RESULT_SCORE_THRESH=1e-4 PREPROC.TRAIN_SHORT_EDGE_SIZE=[640,800] TRAIN.LR_SCHEDULE=3x
R101-FPN-GN (From Scratch)	47.7;41.7 ⬇️ 3	47.4;40.5	28h (on 64 V100s)	9x+GN+Cascade+TrainAug FPN.CASCADE=True BACKBONE.RESNET_NUM_BLOCKS=[3,4,23,3] FPN.NORM=GN BACKBONE.NORM=GN FPN.FRCNN_HEAD_FUNC=fastrcnn_4conv1fc_gn_head FPN.MRCNN_HEAD_FUNC=maskrcnn_up4conv_gn_head PREPROC.TRAIN_SHORT_EDGE_SIZE=[640,800] TRAIN.LR_SCHEDULE=9x BACKBONE.FREEZE_AT=0

1: Numbers taken from Detectron Model Zoo. We compare models that have identical training & inference cost between the two implementations. Their numbers can be different due to small implementation details.

2: Our mAP is 7 point better than the official model in matterport/Mask_RCNN which has the same architecture. Our implementation is also 5x faster.

3: This entry does not use ImageNet pre-training. Detectron numbers are taken from Fig. 5 in Rethinking ImageNet Pre-training. Note that our training strategy is slightly different: we enable cascade throughout the entire training. As far as I know, this model is the best open source TF model on COCO dataset.

Other Datasets / Implementation Details / Speed:

See BALLOON.md and NOTES.md for more details.