/tf-faster-rcnn

Tensorflow Faster RCNN

Primary LanguagePythonMIT LicenseMIT

tf-faster-rcnn

A Tensorflow implementation of faster RCNN detection framework by Xinlei Chen (xinleic@cs.cmu.edu). This 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 and Resnet V1 models. We tested it on plain VGG16 and Resnet101 (thank you @philokey!) architecture so far. 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, 71.2.
  • Train on VOC 2007+2012 trainval and test on VOC 2007 test (R-FCN schedule), 75.3.
  • Train on COCO 2014 trainval35k and test on minival (900k/1190k), 29.5.

With Resnet101 (last conv4):

  • Train on VOC 2007 trainval and test on VOC 2007 test, 75.2.
  • Train on VOC 2007+2012 trainval and test on VOC 2007 test (R-FCN schedule), 79.3.
  • Train on COCO 2014 trainval35k and test on minival (900k/1190k), 34.1.

More Resnets:

  • (Deprecated) Train on COCO 2014 trainval35k and test on minival with approximate FPN baseline setup (900k/1290k), 35.8.

Note:

  • 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 Resnet101, we fix the first block (total 4) when fine-tuning the network, and only use crop_and_resize to resize the RoIs (7x7) without max-pool. 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 Resnet101 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 Tensorflow installation. The code follows r1.0 format. If you are using an order version (r0.1-r0.12), please check out the v0.12 release. While it is not required, for experimenting the original RoI pooling (which requires modification of the C++ code in tensorflow), you can check out my tensorflow fork and look for tf.image.roi_pooling.
  • Python packages you might not have: cython, opencv-python, easydict (similar to py-faster-rcnn). For easydict make sure you have the right version. I use 1.6.
  • Docker users: A Docker image containing all of the required dependencies can be found in Docker hub at the docker folder. The Docker file used to create this image can be found in the docker directory of this repository.

Installation

  1. Clone the repository
git clone https://github.com/endernewton/tf-faster-rcnn.git
  1. Update your -arch in setup script to match your GPU
cd tf-faster-rcnn/lib
vim setup.py
# Check the GPU architecture, if you are using Pascal arch, please switch to sm_61
  1. Build the Cython modules
make clean
make
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 ..

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
# Resnet101 for voc pre-trained on 07+12 set
./data/scripts/fetch_faster_rcnn_models.sh

Note: if you cannot download the models through the link, or you want to try more models, you can check out the following solutions and optionally update the downloading script:

  • 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

Note: If you cannot get the reported numbers (for this one the default mAP is 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 slim, you can get the pre-trained models here 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
wget -v http://download.tensorflow.org/models/vgg_16_2016_08_28.tar.gz
tar -xzvf vgg_16_2016_08_28.tar.gz
mv vgg_16.ckpt vgg16.ckpt
cd ../..

    For Resnet101, you can set up like:

    mkdir -p data/imagenet_weights
cd data/imagenet_weights
wget -v http://download.tensorflow.org/models/resnet_v1_101_2016_08_28.tar.gz
tar -xzvf resnet_v1_101_2016_08_28.tar.gz
mv resnet_v1_101.ckpt res101.ckpt
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: double check you have deleted softlink to the pre-trained models before training!

  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 1. 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

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