Implementation of multi-class object detection using RDFCN for NVIDIA AICity Challenge 2017 and AVSS2017 UA_DETRAC Challenge
This repository contains the source codes of RDFCN implementation for the detection tasks of NVIDIA AICity Challenge and AVSS2017 UA_DETRAC Challenge. This source code is built upon the original Deformable-ConvNets by rearchers from Microsoft Research Asia. The main contribution of this repo to the original repo includes:
-
Add two more data readers: one for NVIDIA AICity Challenge 2017; one for AVSS2017 UA_DETRAC Challenge.
-
Implement a transfer learning
-
Add data inference function to output detection .txt files, one per image.
The codes have been tested on:
-
Ubuntu 16.04, NVIDIA TITIAN X GPU, Python 2.7 (both training and inference).
-
NVIDIA Jetson TX2, Python 2.7 (inference only).
The track one of AI City Challenge by IEEE Smart World and NVIDIA required each paricipating team to detect 14-class objects from urban intersection surveillance cameras using deep learning methods.
Detailed information of NVIDIA AICity Challenge 2017 can be found here.
The detection task of AVSS2017 Challenge required each paricipating team to detect vehicles from UA-DETRAC, a real-world multi-object detection and multi-object tracking benchmark.
Detailed information of AVSS2017 Challenge can be found here.
UA-DETRAC dataset consists of 10 hours of videos captured with a Cannon EOS 550D camera at 24 different locations at Beijing and Tianjin in China.
Detailed information of UA-DETRAC can be found here.
RDFCN stands for Region-based Deformable Fully Convolutional Networks. RDFCN consitas of two major components:
-
R-FCN (Region-based Fully Convolutional Networks). R-FCN is initially described in a NIPS 2016 paper. The Key idea of R-FCN is increasing model inference speed by using more shared convolution.
-
Deformable ConvNets (Deformable Convolutional Networks). Deformable ConvNets is initially described in an arxiv tech report. The key idea of Deformable ConvNets is increasing classification accuracy by using adaptive-shaped convolutional filter.
- Carefully follow the instructions of the offical implementation for Deformable Convolutional Networks based on MXNet here.
At the end of this step:
-
The Deformable ConvNets repo should have been downloaded;
-
MXNet should have been downloaded and properly compiled;
-
The R-FCN demo should be able to run by
python ./rfcn/demo.py.
Note:
For installing MXnet on NVIDIA Jetson TX2, if the above installation instruction doesn't work, you can try following these steps (The MXnet on Jetson TX2 installation has been verified by Koray):
a. MxNet is only compatible with opencv-python >= 3.2.0. Please first upgrade the OpenCV libraries for Jetson platform as described in here
b. Then install MXnet libraries for Jetson TX2 as described here (Devices - NVIDIA Jetson TX2).
- clone this repo into the root/parent directory of Deformable-ConvNets:
cd path/to/Deformable-ConvNets/..
git clone https://github.com/wkelongws/RDFCN_UADETRAC_AICITY
- implement the contribution codes in this repo to the Deformable-ConvNets folder:
cd RDFCN_UADETRAC_AICITY
python setup.py
-
Finish the Installation
-
Download the model weights from here. Put the downloaded file "rfcn_AICityVOC1080_FreezeCOCO_rpnOnly_all_withsignal-0004.params" in the root folder of this repo "RDFCN_UADETRAC_AICITY"
-
Install the provided test images from aic1080 and the downloaded weights (create proper paths and move files accordingly)
cd RDFCN_UADETRAC_AICITY
./test_data_install.sh
- Modify the inference code (optional):
-
Comment line 242 in Deformable-ConvNets/rfcn/functions/inference_rcnn.py if you don't want to see labeled images on run time. By commenting line 242, the program will only output .txt detection files to "Deformable-ConvNets/data/output".
-
Change the threshold at line 234 in Deformable-ConvNets/rfcn/functions/inference_rcnn.py. The current theshold is set at 0.2.
- Test the model by running:
cd Deformable-ConvNets
sh demo.sh
or alternatively:
cd Deformable-ConvNets
python experiments/rfcn/rfcn_Inference_Shuo_AICity.py --cfg experiments/rfcn/cfgs/resnet_v1_101_voc0712_rfcn_dcn_Shuo_AICityVOC1080_FreezeCOCO_rpnOnly_all_withsignal.yaml
-
Download data from UADETRAC and AICity into
data/folder -
Create configuration file
-
cd path/to/Deformable-ConvNets -
Model training and Inference
- To train model from scratch (for both AICITY and UADETRAC), run:
python experiments/rfcn/rfcn_end2end_train_Shuo_UADETRAC.py --cfg path/to/your/configuration/file
Note: For AICITY, based on our experience "training from scratch" doesn't yield good model. Please refer to "Transfer Learning" section for other solution.
- To detect vechiles from the test dataset, run:
python experiments/rfcn/rfcn_Inference_Shuo_UADETRAC.py --cfg path/to/your/configuration/file
python experiments/rfcn/rfcn_Inference_Shuo_AICity.py --cfg path/to/your/configuration/file
Two sample configuration files, one for UADETRAC and one for AICity, have been added to experiments/rfcn/cfgs/
The weights of submitted model (and the corresponding configuration file) on aic1080 can be downloaded here.
The weights of submitted model (and the corresponding configuration file) on aic540 can be downloaded here.
The weights of submitted model (and the corresponding configuration file) on aic480 can be downloaded here.
The weights of submitted model (and the corresponding configuration file) on UADETRAC can be downloaded here.
One major contribution of this repo is adding a transfer learning module to the original implementation.
To train model on AICity using transfer learning, run:
python experiments/rfcn/rfcn_transfer_learning_train_Shuo_AICity.py --cfg path/to/your/configuration/file
There are multiple ways of transfer learning in this RFCN structure:
-
Initialize using pretrained weights then fine-tune all the weight by training on AICITY data.
-
Initialize using pretrained weights, freeze the weights in backbone resnet-101 and RPN, then fine-tune the location-sensitive score maps feature extractor and the output layer by training on AICITY data.
-
Initialize using pretrained weights, freeze all the weights except the outlayer, then only fine-tune the outlayer by training on AICITY data.
Those three strategies have been implemented in "transfer_learning_end2end_freezeNo.py", "transfer_learning_end2end_freezeRPN.py", "transfer_learning_end2end_freezeBoth.py", respectively. To switch between different strategies, replace the codes in "transfer_learning_end2end.py" (placed in rfcn/ after running setup.py) with the one you want. You can choose the initialization weights by changing the codes starting at line 103 in "transfer_learning_end2end.py"
Some insights of how we eventually selected transfer learning (the second strategy) can be found in the presentation "experiments_before_workshop.pptx" in presentations/
Sample detections: (Annotations on the left with red boxes, detections on the right with green boxes, thresholded by condidence 0.1)
Precision-Recall cuve on validation dataset (aic1080):
AICity submission results (ranked 2nd place in all three dataset by 3:08PM Aug 9th, 2017)
The detailed results have been pusblished on NVIDIA AICity challenge website.
- aic480
- aic540
- aic1080
A preprocessed video demonstrating the detection on aic1080 can be found here. 0.2 is used as the detection confidence threshold in the demonstration video.
The workshop presentation can be found in presentations/
Sample detections: (detection with green boxes, thresholded by condidence 0.5)
Precision-Recall cuve on training dataset:
This repo is still partially under contruction. The path of certain files may need correction. For any questions you can contact Shuo Wang.