By Jimmy Ren, Xiaohao Chen, Jianbo Liu, Wenxiu Sun, Jiahao Pang, Qiong Yan, Yu-Wing Tai, Li Xu.
High localization accuracy is crucial in many real-world applications. We propose a novel single stage end-to-end object detection network (RRC) to produce high accuracy detection results. You can use the code to train/evaluate a network for object detection task. For more details, please refer to our paper (https://arxiv.org/abs/1704.05776).
method | KITTI test mAP car (moderate) |
---|---|
Mono3D | 88.66% |
SDP+RPN | 88.85% |
MS-CNN | 89.02% |
Sub-CNN | 89.04% |
RRC (single model) | 89.85% |
Please cite RRC in your publications if it helps your research:
@inproceedings{Ren17CVPR,
author = {Jimmy Ren and Xiaohao Chen and Jianbo Liu and Wenxiu Sun and Jiahao Pang and Qiong Yan and Yu-Wing Tai and Li Xu},
title = {Accurate Single Stage Detector Using Recurrent Rolling Convolution},
booktitle = {CVPR},
year = {2017}
}
- Get the code. We will call the directory that you cloned Caffe into
$CAFFE_ROOT
https://github.com/xiaohaoChen/rrc_detection.git cd rrc_detection
- Build the code. Please follow Caffe instruction to install all necessary packages and build it.
Before build it, you should install CUDA and CUDNN(v5.0).
CUDA 7.5 and CUDNN v5.0 were adapted in our computer.# Modify Makefile.config according to your Caffe installation. cp Makefile.config.example Makefile.config make -j8 # Make sure to include $CAFFE_ROOT/python to your PYTHONPATH. make py make test -j8 make runtest -j8
-
Download fully convolutional reduced (atrous) VGGNet. By default, we assume the model is stored in
$CAFFE_ROOT/models/VGGNet/
. -
Download the KITTI dataset(https://www.cvlibs.net/datasets/kitti/eval_object.php). By default, we assume the data is stored in
$HOME/data/KITTI/
Unzip the training images, testing images and the labels in$HOME/data/KITTI/
. -
Create the LMDB file. For training . As only the images contain cars are adopted as training set for car detection, the labels for cars should be extracted.
We have provided the list of images contain cars in$CAFFE_ROOT/data/KITTI-car/
.# extract the labels for cars cd $CAFFE_ROOT/data/KITTI-car/ ./extract_car_label.sh
Before create the LMDB files. The labels should be converted to VOC type. We provide some matlab scripts.
The scripts are in$CAFFE_ROOT/data/convert_labels/
. Just modifyconverlabels.m
.line 4: root_dir = '/your/path/to/KITTI/';
VOC type labels will be generated in
$KITTI_ROOT/training/labels_2car/xml/
.cd $CAFFE_ROOT/data/KITTI-car/ # Create the trainval.txt, test.txt, and test_name_size.txt in data/KITTI-car/ ./create_list.sh # You can modify the parameters in create_data.sh if needed. # It will create lmdb files for trainval and test with encoded original image: # - $HOME/data/KITTI/lmdb/KITTI-car_training_lmdb/ # - $HOME/data/KITTI/lmdb/KITTI-car_testing_lmdb/ # and make soft links at data/KITTI-car/lmdb ./data/KITTI-car/create_data.sh
- Train your model and evaluate the model.
We train our models in a computer with 4 TITAN X(Maxwell) GPU cards. By default, we assume you train the models on mechines with 4 TITAN X GPUs.
# It will create model definition files and save snapshot models in: # - $CAFFE_ROOT/models/VGGNet/KITTI/RRC_2560x768_kitti_car/ # and job file, log file in: # - $CAFFE_ROOT/jobs/VGGNet/KITIIT/RRC_2560x768_kitti_car/ # After 60k iterations, we can get the model as we said in the paper (mAP 89.*% in KITTI). python examples/car/rrc_kitti_car.py # Before run the testing script. You should modify [line 10: img_dir] to [your path to kitti testing images]. python examples/car/rrc_test.py
If you only have one TITAN X card, you should modify the scriptrrc_kitti.py
.If you have two TITAN X cards, you should modify the scriptline 118: gpus = "0,1,2,3" -> gpus = "0" line 123: batch_size = 4 -> batch_size = 1
rrc_kitti.py
as follow.You can submit the result at kitti submit. If you don't have time to train your model, you can download a pre-trained model from the link as follow.line 118: gpus = "0,1,2,3" -> gpus = "0,1" line 123: batch_size = 4 -> batch_size = 2
Google Drive
Baidu Cloud
Unzip the files in$caffe_root/models/VGGNet/KITTI/
, and run the testing scriptrrc_test.py
, you will get the same result as the single model result we showed in the paper.# before run the script, you should modify the kitti_root at line 10. # Make sure that the work directory is caffe_root cd $caffe_root python models/VGGNet/KITTI/RRC_2560x768_kitti_4r4b_max_size/rrc_test.py
- Evaluate the most recent snapshot.
For testing a model you trained, you show modify the path in
rrc_test.py
.
Thanks to Wei Liu, we have benifited a lot from his previous work SSD (Single Shot Multibox Detector) and his code.