/pytorch-faster-rcnn

Primary LanguagePythonMIT LicenseMIT

pytorch-faster-rcnn

A pytorch implementation of faster RCNN detection framework based on Xinlei Chen's tf-faster-rcnn. Xinlei Chen's repository is based on the python Caffe implementation of faster RCNN available here.

Note: Several minor modifications are made when reimplementing the framework, which give potential improvements. For details about the modifications and ablative analysis, please refer to the technical report An Implementation of Faster RCNN with Study for Region Sampling. If you are seeking to reproduce the results in the original paper, please use the official code or maybe the semi-official code. For details about the faster RCNN architecture please refer to the paper Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks.

Detection Performance

The current code supports VGG16, Resnet V1 and Mobilenet V1 models. We mainly tested it on plain VGG16 and Resnet101 architecture. As the baseline, we report numbers using a single model on a single convolution layer, so no multi-scale, no multi-stage bounding box regression, no skip-connection, no extra input is used. The only data augmentation technique is left-right flipping during training following the original Faster RCNN. All models are released.

With VGG16 (conv5_3):

  • Train on VOC 2007 trainval and test on VOC 2007 test, 70.48(crop and resize), 69.95(roi pooling) (71.2 for tf-faster-rcnn).
  • Train on VOC 2007+2012 trainval and test on VOC 2007 test (R-FCN schedule), 74.83(crop and resize) 74.59(roi pooling)(75.3 for tf-faster-rcnn).
  • Train on COCO 2014 trainval35k and test on minival (900k/1190k) 27.0 (29.5 for tf-faster-rcnn). ((350k/490k) 24.6(crop and resize) 21.8(roi pooling)).

With Resnet101 (last conv4):

  • Train on VOC 2007 trainval and test on VOC 2007 test, 73.72(crop and resize) 75.07(roi pooling) (75.2 for tf-faster-rcnn).
  • Train on VOC 2007+2012 trainval and test on VOC 2007 test (R-FCN schedule), 78.28(crop and resize) 79.57(roi pooling) (79.3 for tf-faster-rcnn).
  • Train on COCO 2014 trainval35k and test on minival (800k/1190k), 34.4 （34.1 for tf-faster-rcnn). ((350k/490k) 30.3(crop and resize) 28.2(roi pooling)).

More Results:

  • Train Mobilenet (1.0, 224) on COCO 2014 trainval35k and test on minival (900k/1190k), 21.9.
  • Train Resnet50 on COCO 2014 trainval35k and test on minival (900k/1190k), 31.6.
  • Train Resnet152 on COCO 2014 trainval35k and test on minival (900k/1190k), 35.2.

Approximate baseline setup from FPN (this repo does not contain training code for FPN yet):

  • Train Resnet50 on COCO 2014 trainval35k and test on minival (900k/1190k), 33.4.
  • Train Resnet101 on COCO 2014 trainval35k and test on minival (900k/1190k), 36.3.
  • Train Resnet152 on COCO 2014 trainval35k and test on minival (1000k/1390k), 37.2.

Note:

  • Compared to tf-faster-rcnn, we use roi pooling instead of crop_and_resize; we don't know how this affects result compared to tf-faster-rcnn.
  • ~~Due to the randomness in GPU training with Tensorflow espeicially for VOC, the best numbers are reported (with 2-3 attempts) here. According to my experience, for COCO you can almost always get a very close number (within ~0.2%) despite the randomness.~~
  • All the numbers are obtained with a different testing scheme without selecting region proposals using non-maximal suppression (TEST.MODE top), the default and original testing scheme (TEST.MODE nms) will likely result in slightly worse performance (see report, for COCO it drops 0.X AP).
  • Since we keep the small proposals (< 16 pixels width/height), our performance is especially good for small objects.
  • For other minor modifications, please check the report. Notable ones include using crop_and_resize, and excluding ground truth boxes in RoIs during training.
  • For COCO, we find the performance improving with more iterations (VGG16 350k/490k: 26.9, 600k/790k: 28.3, 900k/1190k: 29.5), and potentially better performance can be achieved with even more iterations.
  • For Resnets, we fix the first block (total 4) when fine-tuning the network, and only use crop_and_resize roi pooling to resize the RoIs (7x7) without max-pool (which I find useless especially for COCO). The final feature maps are average-pooled for classification and regression. All batch normalization parameters are fixed. Weight decay is set to Renset101 default 1e-4. Learning rate for biases is not doubled.
  • For approximate FPN baseline setup we simply resize the image with 800 pixels, add 32^2 anchors, and take 1000 proposals during testing.
  • Check out here/here/here for the latest models, including longer COCO VGG16 models and Resnet ones.

Additional features

Additional features not mentioned in the report are added to make research life easier:

  • Support for train-and-validation. During training, the validation data will also be tested from time to time to monitor the process and check potential overfitting. Ideally training and validation should be separate, where the model is loaded everytime to test on validation. However I have implemented it in a joint way to save time and GPU memory. Though in the default setup the testing data is used for validation, no special attempts is made to overfit on testing set.
  • Support for resuming training. I tried to store as much information as possible when snapshoting, with the purpose to resume training from the lateset snapshot properly. The meta information includes current image index, permutation of images, and random state of numpy. However, when you resume training the random seed for tensorflow will be reset (not sure how to save the random state of tensorflow now), so it will result in a difference. Note that, the current implementation still cannot force the model to behave deterministically even with the random seeds set. Suggestion/solution is welcome and much appreciated.
  • Support for visualization. The current implementation will summarize ground truth detections, statistics of losses, activations and variables during training, and dump it to a separate folder for tensorboard visualization. The computing graph is also saved for debugging.

Prerequisites

  • A basic pytorch installation. The code follows 0.2. If you are using old 0.1.12, you can checkout 0.1.12 branch.
  • Python packages you might not have: cffi, opencv-python, easydict (similar to py-faster-rcnn). For easydict make sure you have the right version. I use 1.6.
  • tensorboard-pytorch to visualize the training and validation curve. Please build from source to use the latest tensorflow-tensorboard.
  • Docker users: Since the recent upgrade, the docker image on docker hub (https://hub.docker.com/r/mbuckler/tf-faster-rcnn-deps/) is no longer valid. However, you can still build your own image by using dockerfile located at docker folder (cuda 8 version, as it is required by Tensorflow r1.0.) And make sure following Tensorflow installation to install and use nvidia-docker[https://github.com/NVIDIA/nvidia-docker]. Last, after launching the container, you have to build the Cython modules within the running container.

Installation

  1. Clone the repository
git clone https://github.com/ruotianluo/pytorch-faster-rcnn.git
  1. Update your -arch in setup script to match your GPU
cd pytorch-faster-rcnn/lib
# Change the GPU architecture (-arch) if necessary
vim setup.py
GPU model Architecture
TitanX (Maxwell/Pascal) sm_52
Grid K520 (AWS g2.2xlarge) sm_30
Tesla K80 (AWS p2.xlarge) sm_37

Note: You are welcome to contribute the settings on your end if you have made the code work properly on other GPUs.

  1. Build RoiPooling module
cd lib/layer_utils/roi_pooling/src/cuda
echo "Compiling roi_pooling kernels by nvcc..."
nvcc -c -o roi_pooling_kernel.cu.o roi_pooling_kernel.cu -x cu -Xcompiler -fPIC -arch=sm_52
cd ../../
python build.py
cd ../../../
  1. Build NMS
cd lib/nms/src/cuda
echo "Compiling nms kernels by nvcc..."
nvcc -c -o nms_kernel.cu.o nms_kernel.cu -x cu -Xcompiler -fPIC -arch=sm_52
cd ../../
python build.py
cd ../../
  1. Install the Python COCO API. The code requires the API to access COCO dataset.
cd data
git clone https://github.com/pdollar/coco.git
cd coco/PythonAPI
make
cd ../../..

Setup data

Please follow the instructions of py-faster-rcnn here to setup VOC and COCO datasets (Part of COCO is done). The steps involve downloading data and optionally creating softlinks in the data folder. Since faster RCNN does not rely on pre-computed proposals, it is safe to ignore the steps that setup proposals.

If you find it useful, the data/cache folder created on my side is also shared here.

Demo and Test with pre-trained models

  1. Download pre-trained model (only google drive works)
  • Another server here.
  • Google drive here.
  1. Create a folder and a softlink to use the pre-trained model
NET=res101
TRAIN_IMDB=voc_2007_trainval+voc_2012_trainval
mkdir -p output/${NET}/${TRAIN_IMDB}
cd output/${NET}/${TRAIN_IMDB}
ln -s ../../../data/voc_2007_trainval+voc_2012_trainval ./default
cd ../../..
  1. Demo for testing on custom images
# at reposistory root
GPU_ID=0
CUDA_VISIBLE_DEVICES=${GPU_ID} ./tools/demo.py

Note: Resnet101 testing probably requires several gigabytes of memory, so if you encounter memory capacity issues, please install it with CPU support only. Refer to Issue 25.

  1. Test with pre-trained Resnet101 models
GPU_ID=0
./experiments/scripts/test_faster_rcnn.sh $GPU_ID pascal_voc_0712 res101

Note: If you cannot get the reported numbers (78.7 on my side), then probabaly the NMS function is compiled improperly, refer to Issue 5.

Train your own model

  1. Download pre-trained models and weights. The current code support VGG16 and Resnet V1 models. Pre-trained models are provided by pytorch-vgg and pytorch-resnet (the ones with caffe in the name), you can download the pre-trained models and set them in the data/imagenet_weights folder. For example for VGG16 model, you can set up like:

    mkdir -p data/imagenet_weights
cd data/imagenet_weights
python # open python in terminal and run the following Python code
    import torch
from torch.utils.model_zoo import load_url
from torchvision import models

sd = load_url("https://s3-us-west-2.amazonaws.com/jcjohns-models/vgg16-00b39a1b.pth")
sd['classifier.0.weight'] = sd['classifier.1.weight']
sd['classifier.0.bias'] = sd['classifier.1.bias']
del sd['classifier.1.weight']
del sd['classifier.1.bias']

sd['classifier.3.weight'] = sd['classifier.4.weight']
sd['classifier.3.bias'] = sd['classifier.4.bias']
del sd['classifier.4.weight']
del sd['classifier.4.bias']

torch.save(sd, "vgg16.pth")
    cd ../..

    For Resnet101, you can set up like:

    mkdir -p data/imagenet_weights
cd data/imagenet_weights
# download from my gdrive (link in pytorch-resnet)
mv resnet101-caffe.pth res101.pth
cd ../..

  2. Train (and test, evaluation)

./experiments/scripts/train_faster_rcnn.sh [GPU_ID] [DATASET] [NET]
# GPU_ID is the GPU you want to test on
# NET in {vgg16, res50, res101, res152} is the network arch to use
# DATASET {pascal_voc, pascal_voc_0712, coco} is defined in train_faster_rcnn.sh
# Examples:
./experiments/scripts/train_faster_rcnn.sh 0 pascal_voc vgg16
./experiments/scripts/train_faster_rcnn.sh 1 coco res101

Note: Please double check you have deleted softlink to the pre-trained models before training. If you find NaNs during training, please refer to Issue 86.

  1. Visualization with Tensorboard
tensorboard --logdir=tensorboard/vgg16/voc_2007_trainval/ --port=7001 &
tensorboard --logdir=tensorboard/vgg16/coco_2014_train+coco_2014_valminusminival/ --port=7002 &
  1. Test and evaluate
./experiments/scripts/test_faster_rcnn.sh [GPU_ID] [DATASET] [NET]
# GPU_ID is the GPU you want to test on
# NET in {vgg16, res50, res101, res152} is the network arch to use
# DATASET {pascal_voc, pascal_voc_0712, coco} is defined in test_faster_rcnn.sh
# Examples:
./experiments/scripts/test_faster_rcnn.sh 0 pascal_voc vgg16
./experiments/scripts/test_faster_rcnn.sh 1 coco res101
  1. You can use tools/reval.sh for re-evaluation

By default, trained networks are saved under:

output/[NET]/[DATASET]/default/

Test outputs are saved under:

output/[NET]/[DATASET]/default/[SNAPSHOT]/

Tensorboard information for train and validation is saved under:

tensorboard/[NET]/[DATASET]/default/
tensorboard/[NET]/[DATASET]/default_val/

The default number of training iterations is kept the same to the original faster RCNN for VOC 2007, however I find it is beneficial to train longer (see report for COCO), probably due to the fact that the image batch size is one. For VOC 07+12 we switch to a 80k/110k schedule following R-FCN. Also note that due to the nondeterministic nature of the current implementation, the performance can vary a bit, but in general it should be within ~1% of the reported numbers for VOC, and ~0.2% of the reported numbers for COCO. Suggestions/Contributions are welcome.

Citation

If you find this implementation or the analysis conducted in our report helpful, please consider citing:

@article{chen17implementation,
    Author = {Xinlei Chen and Abhinav Gupta},
    Title = {An Implementation of Faster RCNN with Study for Region Sampling},
    Journal = {arXiv preprint arXiv:1702.02138},
    Year = {2017}
}

For convenience, here is the faster RCNN citation:

@inproceedings{renNIPS15fasterrcnn,
    Author = {Shaoqing Ren and Kaiming He and Ross Girshick and Jian Sun},
    Title = {Faster {R-CNN}: Towards Real-Time Object Detection
             with Region Proposal Networks},
    Booktitle = {Advances in Neural Information Processing Systems ({NIPS})},
    Year = {2015}
}

Detailed numbers from COCO server (not supported)

All the models are trained on COCO 2014 trainval35k.

VGG16 COCO 2015 test-dev (900k/1190k):

 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.297
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.504
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.312
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.128
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.325
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.421
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.272
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.399
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.409
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.187
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.451
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.591

VGG16 COCO 2015 test-std (900k/1190k):

Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.295
Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 0.501
Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.312
Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.119
Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.327
Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.418
Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.273
Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.400
Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.409
Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.179
Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.455
Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.586