Official Tensorflow implementation of drl-RPN by Aleksis Pirinen (email: [email protected]) and Cristian Sminchisescu. The associated CVPR 2018 paper can be accessed here. A video demonstrating this work can be seen here.
The drl-RPN model is implemented on top of the publicly available TensorFlow VGG-16-based Faster R-CNN implementation by Xinlei Chen available here. See also the associated technical report An Implementation of Faster RCNN with Study for Region Sampling, as well as the original Faster R-CNN paper Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks.
- A basic Tensorflow installation. The code follows r1.2 format.
- Python packages you might not have: cython, opencv-python, easydict (similar to py-faster-rcnn). For easydict make sure you have the right version (1.6 was used here).
- See also "Prerequisites" on this page.
- Clone the repository
git clone https://github.com/aleksispi/drl-rpn-tf.git
- For steps 2-4, see "Installation" on this page.
The current code supports VGG16 models. Exactly as for the Faster R-CNN implementation by Xinlei Chen, 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 R-CNN.
We first re-ran some of the experiments reported here for Faster R-CNN, but trained the models longer to obtain further performance gains for our baseline models. We got:
- Train on VOC 2007+2012 trainval (iterations: 100k/180k) and test on VOC 2007 test (trained like here, but for more iterations), 76.5.
- Train on VOC 2007+2012 trainval + 2007 test (iterations: 100k/180k) and test on VOC 2012 test, 74.2.
The corresponding results when using our drl-RPN detector with exploration penalty 0.05 during inference (models trained over different exploration penalties, as described in Section 5.1.2 in the paper) and posterior class-probability adjustments (Section 4.2 in our paper):
- Train on VOC 2007+2012 trainval (iterations: 90k/110k for core model, 80k/110k for posterior class-probability adjustment module) and test on VOC 2007 test (trained like here, but for more iterations), 77.5. Without posterior class-probability adjustments (np): 77.2. Average exploration (% RoIs forwarded per image on average): 28.0%. Average number of fixations per image: 5.6.
- Train on VOC 2007+2012 trainval + 2007 test (iterations: 90k/110k, 80k/110k for posterior class-probability adjustment module) and test on VOC 2012 test, 74.9. Without posterior class-probability adjustments (np): 74.6. Average exploration (% RoIs forwarded per image on average): 30.6%. Average number of fixations per image: 6.7.
Tabular result representation
Model | mAP - VOC 2007 | mAP - VOC 2012 |
---|---|---|
RPN | 76.5 | 74.2 |
drl-RPN | 77.5 | 74.9 |
drl-RPN (np) | 77.2 | 74.6 |
drl-RPN (12-fix) | 77.6 | 75.0 |
Note:
- All settings are shared with that of Xinlei Chen for the things relating to the baseline Faster R-CNN model (RPN).
- See the code for any deviations from the CVPR 2018 paper. Some important changes post-CVPR:
- Training over different exploration-accuracy trade-offs is now the default model (as opposed to training for a fixed exploration penalty). Hence the default model allows for setting the exploration-accuracy trade-off during testing (c.f. Section 5.1.2 and Figure 6 in the paper). Turns out we only need two different exploration penalties (0.05 and 0.35 was used), but setting any other trade-off parameters during inference is possible.
- Separation of rewards (Section 5.1.1 in the paper) does not yield accuracy gains for models trained over different exploration-accuracy trade-offs, so it is not used. See
reward_functions.py
for details. - The drl-RPN models are now much more fast to train than how it was done in the original paper (c.f. Section 5.2). Specifically, instead of sampling 50 search trajectories per image to estimate the policy gradient, we now run 50 search trajectories on 50 different images. This reduces training time by 5-10 times, yet we get results in the same ball park.
All pretrained models (both Faster R-CNN baseline and our drl-RPN models) for the numbers reported above in Detection Performance are available:
- drl-RPN trained on VOC 2007+2012 trainval: https://www.dropbox.com/s/jjioouboj2fytjz/drl-rpn-voc2007-2012-trainval.tar.gz?dl=0
- drl-RPN trained on VOC 2007+2012 trainval + 2007 test: https://www.dropbox.com/s/3srmz5z4qtpntyp/drl-rpn-voc2007-2012-trainval-plus-2007test.tar.gz?dl=0
- Faster R-CNN trained on VOC 2007+2012 trainval: https://www.dropbox.com/s/5bvlrnc3mmpbb6a/fr-rcnn-voc2007-2012-trainval.tar.gz?dl=0
- Faster R-CNN trained on VOC 2007+2012 trainval + 2007 test: https://www.dropbox.com/s/y64w4h81due9vcu/fr-rcnn-voc2007-2012-trainval%2B2007test.tar.gz?dl=0
See "Setup data" on this page. Essentially download the dataset you are interested (e.g. PASCAL VOC), and add soft links in the data
folder in the appropriate way (see https://askubuntu.com/questions/56339/how-to-create-a-soft-or-symbolic-link for generic how-to for setting soft links).
- Download and setup the datasets (see Object detection datasets above).
- Download the desired pretrained Faster R-CNN model (see Pretrained models above).
- The main script to launch training is
experiments/scripts/train_drl_rpn.sh
. SetupSAVE_PATH
andWEIGHT_PATH
appropriately, and run the command./experiments/scripts/train_drl_rpn.sh 0 pascal_voc_0712 1 20000 0 110000
to start training on VOC 2007+2012 trainval on GPU-id 0 for a total of 110k iterations (see code for more details). This will yield a drl-RPN model trained over two exploration penalties, enabling setting the speed-accuracy trade-off at test time. See alsoexperiments/cfgs/drl-rpn-vgg16.yml
for some settings. - Once the above model has finished training in step 3, it is also possible to train the posterior class-probability history module (c.f. Section 4.2 in our paper). To do this, first make sure that the
WEIGHTS_PATH
variable intrain_drl_rpn.sh
points to your drl-RPN model weights obtained in step 3 above. Then run./experiments/scripts/train_drl_rpn.sh 0 pascal_voc_0712 1 0 1 110000
to train the posterior class-probability adjustment module for 110k iterations.
- Either make sure you have trained your own drl-RPN model (see Training drl-RPN above) or download pretrained weights (see Pretrained models above).
- The main script to launch testing is
experiments/scripts/test_drl_rpn.sh
. To test your model on the Pascal VOC 2007 test set on GPU-id 0, run./experiments/scripts/test_drl_rpn.sh 0 pascal_voc_0712 1 1 0
(see code for more details). If you want to change the exploration-accuracy trade-off parameter, seeexperiments/cfgs/drl-rpn-vgg16.yml
. You may also specify whether you want to visualize drl-RPN search trajectories here (visualizations are saved in the top folder).
Here are solutions to some potential issues:
- Problem:
import pycocotools._mask as _mask
ImportError: No module named _mask
Command exited with non-zero status 1
- Solution: Go to
data/coco/PythonAPI/
and runmake
in your ubuntu terminal. Now it should be fine!
If you find this implementation or our CVPR 2018 paper interesting or helpful, please consider citing:
@inproceedings{pirinen2018deep,
title={Deep reinforcement learning of region proposal networks for object detection},
author={Pirinen, Aleksis and Sminchisescu, Cristian},
booktitle={proceedings of the IEEE conference on computer vision and pattern recognition},
pages={6945--6954},
year={2018}
}