easy-faster-rcnn.pytorch

An easy implementation of Faster R-CNN in PyTorch.

Demo

1. Download checkpoint from here
2. Follow the instructions in Setup 2 & 3
3. Run inference script

   $ python infer.py -s=coco2017 -b=resnet101 -c=/path/to/checkpoint.pth --image_min_side=800 --image_max_side=1333 --anchor_sizes="[64, 128, 256, 512]" --rpn_post_nms_top_n=1000 /path/to/input/image.jpg /path/to/output/image.jpg
   

Features

• Supports PyTorch 1.0
• Supports PASCAL VOC 2007 and MS COCO 2017 datasets
• Supports ResNet-18, ResNet-50 and ResNet-101 backbones (from official PyTorch model)
• Supports ROI Pooling and ROI Align pooler modes
• Supports Multi-Batch and Multi-GPU training
• Matches the performance reported by the original paper
• It's efficient with maintainable, readable and clean code

Benchmarking

• PASCAL VOC 2007

  • Train: 2007 trainval (5011 images)
  • Eval: 2007 test (4952 images)

  Implementation	Backbone	GPU	#GPUs	#Batches/GPU	Training Speed (FPS)	Inference Speed (FPS)	mAP	image_min_side	image_max_side	anchor_ratios	anchor_sizes	pooler_mode	rpn_pre_nms_top_n (train)	rpn_post_nms_top_n (train)	rpn_pre_nms_top_n (eval)	rpn_post_nms_top_n (eval)	anchor_smooth_l1_loss_beta	proposal_smooth_l1_loss_beta	batch_size	learning_rate	momentum	weight_decay	step_lr_sizes	step_lr_gamma	warm_up_factor	warm_up_num_iters	num_steps_to_finish
  Original Paper	VGG-16	Tesla K40	1	1	-	~ 5	0.699	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
  ruotianluo/pytorch-faster-rcnn	ResNet-101	TITAN Xp	-	-	-	-	0.7576	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
  jwyang/faster-rcnn.pytorch	ResNet-101	TITAN Xp	1	1	-	-	0.752	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
  Ours	ResNet-101	GTX 1080 Ti	1	4	7.12	15.05	0.7562	600	1000	[(1, 2), (1, 1), (2, 1)]	[128, 256, 512]	align	12000	2000	6000	300	1.0	1.0	4	0.004	0.9	0.0005	[12500, 17500]	0.1	0.3333	500	22500

• MS COCO 2017

  • Train: 2017 Train drops images without any objects (117266 images)
  • Eval: 2017 Val drops images without any objects (4952 images)

  Implementation	Backbone	GPU	#GPUs	#Batches/GPU	Training Speed (FPS)	Inference Speed (FPS)	[email protected][.5:.95]	[email protected][.5]	[email protected][.75]	AP S	AP M	AP L	image_min_side	image_max_side	anchor_ratios	anchor_sizes	pooler_mode	rpn_pre_nms_top_n (train)	rpn_post_nms_top_n (train)	rpn_pre_nms_top_n (eval)	rpn_post_nms_top_n (eval)	anchor_smooth_l1_loss_beta	proposal_smooth_l1_loss_beta	batch_size	learning_rate	momentum	weight_decay	step_lr_sizes	step_lr_gamma	warm_up_factor	warm_up_num_iters	num_steps_to_finish
  ruotianluo/pytorch-faster-rcnn	ResNet-101	TITAN Xp	-	-	-	-	0.354	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
  jwyang/faster-rcnn.pytorch	ResNet-101	TITAN Xp	8	2	-	-	0.370	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-	-
  Ours	ResNet-101	GTX 1080 Ti	1	2	4.84	8.00	0.356	0.562	0.389	0.176	0.398	0.511	800	1333	[(1, 2), (1, 1), (2, 1)]	[64, 128, 256, 512]	align	12000	2000	6000	1000	0.1111	1.0	2	0.0025	0.9	0.0001	[480000, 640000]	0.1	0.3333	500	720000
  Ours	ResNet-101	Telsa P100	4	4	11.64	5.10	0.370	0.576	0.403	0.187	0.414	0.522	800	1333	[(1, 2), (1, 1), (2, 1)]	[64, 128, 256, 512]	align	12000	2000	6000	1000	0.1111	1.0	16	0.02	0.9	0.0001	[120000, 160000]	0.1	0.3333	500	180000

• PASCAL VOC 2007 Cat Dog

  • Train: 2007 trainval drops categories other than cat and dog (750 images)
  • Eval: 2007 test drops categories other than cat and dog (728 images)

• MS COCO 2017 Person

  • Train: 2017 Train drops categories other than person (64115 images)
  • Eval: 2017 Val drops categories other than person (2693 images)

• MS COCO 2017 Car

  • Train: 2017 Train drops categories other than car (12251 images)
  • Eval: 2017 Val drops categories other than car (535 images)

• MS COCO 2017 Animal

  • Train: 2017 Train drops categories other than bird, cat, dog, horse, sheep, cow, elephant, bear, zebra and giraffe (23989 images)
  • Eval: 2017 Val drops categories other than bird, cat, dog, horse, sheep, cow, elephant, bear, zebra and giraffe (1016 images)

Requirements

• Python 3.6

• torch 1.0

• torchvision 0.2.1

• tqdm

  $ pip install tqdm
  

• tensorboardX

  $ pip install tensorboardX
  

• OpenCV 3.4 (required by infer_stream.py)

  $ pip install opencv-python~=3.4
  

• websockets (required by infer_websocket.py)

  $ pip install websockets
  

Setup

1. Prepare data

   1. For PASCAL VOC 2007

      1. Download dataset

         • Training / Validation (5011 images)
         • Test (4952 images)

      2. Extract to data folder, now your folder structure should be like:

         easy-faster-rcnn.pytorch
             - data
                 - VOCdevkit
                     - VOC2007
                         - Annotations
                             - 000001.xml
                             - 000002.xml
                             ...
                         - ImageSets
                             - Main
                                 ...
                                 test.txt
                                 ...
                                 trainval.txt
                                 ...
                         - JPEGImages
                             - 000001.jpg
                             - 000002.jpg
                             ...
                 - ...
         

   2. For MS COCO 2017

      1. Download dataset

         • 2017 Train images [18GB] (118287 images)

           COCO 2017 Train = COCO 2015 Train + COCO 2015 Val - COCO 2015 Val Sample 5k

         • 2017 Val images [1GB] (5000 images)

           COCO 2017 Val = COCO 2015 Val Sample 5k (formerly known as minival)

         • 2017 Train/Val annotations [241MB]

      2. Extract to data folder, now your folder structure should be like:

         easy-faster-rcnn.pytorch
             - data
                 - COCO
                     - annotations
                         - instances_train2017.json
                         - instances_val2017.json
                         ...
                     - train2017
                         - 000000000009.jpg
                         - 000000000025.jpg
                         ...
                     - val2017
                         - 000000000139.jpg
                         - 000000000285.jpg
                         ...
                 - ...
         

2. Build Non Maximum Suppression and ROI Align modules (modified from facebookresearch/maskrcnn-benchmark)

   1. Install

      $ python support/setup.py develop
      

   2. Uninstall

      $ python support/setup.py develop --uninstall
      

   3. Test

      $ python test/nms/test_nms.py
      

      • Result

        

3. Install pycocotools for MS COCO 2017 dataset

   1. Clone and build COCO API

      $ git clone https://github.com/cocodataset/cocoapi
      $ cd cocoapi/PythonAPI
      $ make
      

      It's not necessary to be under project directory

   2. If an error with message pycocotools/_mask.c: No such file or directory has occurred, please install cython and try again

      $ pip install cython
      

   3. Copy pycocotools into project

      $ cp -R pycocotools /path/to/project
      

Usage

1. Train

   • To apply default configuration (see also config/)

     $ python train.py -s=voc2007 -b=resnet101
     

   • To apply custom configuration (see also train.py)

     $ python train.py -s=voc2007 -b=resnet101 --weight_decay=0.0001
     

   • To apply recommended configuration (see also scripts/)

     $ bash ./scripts/voc2007/train-bs2.sh resnet101 /path/to/outputs/dir
     

2. Evaluate

   • To apply default configuration (see also config/)

     $ python eval.py -s=voc2007 -b=resnet101 /path/to/checkpoint.pth
     

   • To apply custom configuration (see also eval.py)

     $ python eval.py -s=voc2007 -b=resnet101 --rpn_post_nms_top_n=1000 /path/to/checkpoint.pth
     

   • To apply recommended configuration (see also scripts/)

     $ bash ./scripts/voc2007/eval.sh resnet101 /path/to/checkpoint.pth
     

3. Infer

   • To apply default configuration (see also config/)

     $ python infer.py -s=voc2007 -b=resnet101 -c=/path/to/checkpoint.pth /path/to/input/image.jpg /path/to/output/image.jpg
     

   • To apply custom configuration (see also infer.py)

     $ python infer.py -s=voc2007 -b=resnet101 -c=/path/to/checkpoint.pth -p=0.9 /path/to/input/image.jpg /path/to/output/image.jpg
     

   • To apply recommended configuration (see also scripts/)

     $ bash ./scripts/voc2007/infer.sh resnet101 /path/to/checkpoint.pth /path/to/input/image.jpg /path/to/output/image.jpg
     

4. Infer other sources

   • Source from stream (see also infer_stream.py)

     # Camera
     $ python infer_stream.py -s=voc2007 -b=resnet101 -c=/path/to/checkpoint.pth -p=0.9 0 5
     
     # Video
     $ python infer_stream.py -s=voc2007 -b=resnet101 -c=/path/to/checkpoint.pth -p=0.9 /path/to/file.mp4 5
     
     # Remote
     $ python infer_stream.py -s=voc2007 -b=resnet101 -c=/path/to/checkpoint.pth -p=0.9 rtsp://184.72.239.149/vod/mp4:BigBuckBunny_115k.mov 5
     

   • Source from websocket (see also infer_websocket.py)

     1. Start web server

        $ cd webapp
        $ python -m http.server 8000
        

     2. Launch service

        $ python infer_websocket.py -s=voc2007 -b=resnet101 -c=/path/to/checkpoint.pth -p=0.9
        

     3. Navigate website: http://127.0.0.1:8000/

        

        Sample video from Pexels

Notes

• Illustration for "find labels for each anchor_bboxes" in region_proposal_network.py

  

  

• Illustration for NMS CUDA

  

• Plot of beta smooth L1 loss function