VLASE: Vehicle Localization by Aggregating Semantic Edges
VLASE is a framework to use semantic edge features from images to achieve on-road localization. Semantic edge features denote edge contours that separate pairs of distinct objects such as building-sky, roadsidewalk, and building-ground. While prior work has shown promising results by utilizing the boundary between prominent
classes such as sky and building using skylines, we generalize this approach to consider semantic edge features that arise from 19 different classes. Our localization algorithm is simple, yet very powerful. We extract semantic edge features using a recently introduced CASENet architecture and utilize VLAD framework to perform image retrieval. Our experiments show that we achieve improvement over some of the state-of-the-art localization algorithms such as SIFT-VLAD and its deep variant NetVLAD. We use ablation study to study the importance of different semantic classes, and show that our unified approach achieves better performance compared to individual prominent features such as skylines.
** This code is ONLY for academic purposes. If you are interested in commercialising the code, please contact Srikumar Ramalingam (srikumar@cs.utah.edu)
Extract all the frames from the source dashcam video into a directory (We'll call it src). VLASE uses original frames for SIFT baselines.
If you want to you VLASE with CASENet features then extract the CASENet features for the original frames into a directory (We'll call it result_root). result_root will have a directory structure like following -
These 19 directories contain features of each frame for 19 semantic classes. If there are total N frames then each of these directories will have N files.
VLASE trains and tests on different sets of frames. So split your data into training and testing images.
You must have a training and a testing data file with the GPS information of the frames. These files must have the data in following space separated format:
Frame_Name X Y
AM09_000000.png 36.37261637 127.3641256
AM09_000001.png 36.37261636 127.3641256
Refer to the directory sample_data for sample training and testing files.
X and Y are the gps coordinates of the frames. These can be global latitude and longitude or they can be in meters (relative to the starting frame).
Types of experiments:
Sift
Standard: All features equally weighted
Sift with changing the value of alpha (Assign different weights to sift features and xy coordinate features)
Casenet
Standard: All features equally weighted
Casenet with changing the value of alpha (Assign different weights to casenet features and xy coordinate features)
Casenet with removing a subset of features (All Static, Combination of some static)
Execution Instructions:
Build Vocab: scripts/build_vlad_vocabulary.py
First of all, VLASE builds vocabulary using the training frames. It creates and saves the vocabulary file in the same directory. Following are the input arguments for this script -
Required
'-i', '--input_list', type=str,required=True,help=Absolute path of the training images list with gps information
'-src', '--src', type=str,required=True,help=Directory containing original source training images
Casenet specific arguments
'-r', '--result_root', type=str,default=None,help=Directory containing casenet features of training images
'-t', '--thresh', type=float, default='0.5',help=Probability threshold value to use casenet features
'-rc', '--result_classes', type=int, default=19,help=Number of casenet classes
'-crf', '--removed_class',type=str, default=None,help=Comma seperated list of casenet class ids that should be removed during the experiment
Optional
'-f','--feature_type', type=str, default='sift',help=Feature type used for localization - casenet or sift. By default, the script uses SIFT features.
'-c', '--vlad_feature_dimension', type=int, default='32',help=Number of clusters for VLAD. In our experiments, SIFT performed best with cluster=32 and Casenet was best with cluster=64.
'-a', '--alpha', type=float, default='0.5',help=Alpha gives weight to the features (Casenet/SIFT vs XY). XY will have weight = (1-alpha).alpha=0.5 means equal weight to both. In our experiments, alpha=0.1 gave the best results.
'-s', '--skip_count', type=int, default='0',help=Number of frames to skip while training and testing VLAD. To reduce the execution time, we skip video frames. It did not make much difference in the performance as we skip according to the frame rate of videos. For e.g. KAIST dataset has frame rate = 10, so we skip 5 frames while creating vocabulary as it does not cause much loss of information.
```
Examples -
- Using Casenet with Removing features -
python build_vlad_vocabulary.py -i /home/sagar/KAIST/train_GPS.txt -r /home/sagar/KAIST/KAIST_CASENET/train -src /home/sagar/KAIST/src -f casenet -c 64 -s 5 -crf 11,12,13
- Using Sift Standard
python build_vlad_vocabulary.py -i /home/sagar/KAIST/train_GPS.txt -src /home/sagar/KAIST/src -s 5
After building vocab, VLASE creates the VLAD descriptors for training and testing frames. It creates and saves the vlad codebook file in the same directory. Following are the input arguments for this script -
Required
'-tr', '--train_list', type=str,required=True,help=Absolute path of the training images list with gps information. This is the same list used for building vocabulary.
'-te', '--test_list', type=str,required=True,help=Absolute path of the testing images list with gps information.
'-src', '--src', type=str,required=True,help=Directory containing original source images, this directory must have both training and testing data.
Casenet specific arguments
'-r', '--result_root', type=str,default=None,help=Directory containing casenet features, this directory must have both training and testing data.
'-t', '--thresh', type=float, default='0.5',help=Probability threshold value to use casenet features
'-rc', '--result_classes', type=int, default=19,help=Number of casenet classes
'-crf', '--removed_class',type=str, default=None,help=Comma seperated list of casenet class ids that should be removed during the experiment
Optional
'-f','--feature_type', type=str, default='sift',help=Feature type used for localization - casenet or sift. By default, the script uses SIFT features.
'-c', '--vlad_feature_dimension', type=int, default='32',help=Number of clusters for VLAD. In our experiments, SIFT performed best with cluster=32 and Casenet was best with cluster=64.
'-a', '--alpha', type=float, default='0.5',help=Alpha gives weight to the features (Casenet/SIFT vs XY). XY will have weight = (1-alpha).alpha=0.5 means equal weight to both. In our experiments, alpha=0.1 gave the best results.
'-s', '--skip_count', type=int, default='0',help=Number of frames to skip while training and testing VLAD. To reduce the execution time, we skip video frames. It did not make much difference in the performance as we skip according to the frame rate of videos. For e.g. KAIST dataset has frame rate = 10, so we skip 5 frames while creating vocabulary as it does not cause much loss of information.
```
Examples -
- Using Casenet with Removing features
python compute_vlad_descriptor.py -tr /home/sagar/KAIST/train_GPS.txt -te /home/sagar/KAIST/test_GPS.txt -r /home/sagar/KAIST/KAIST_CASENET/all -src /home/sagar/KAIST/all -f casenet -c 64 -s 5 -crf 11,12,13
- Using Sift Standard
python compute_vlad_descriptor.py -tr /home/sagar/KAIST/train_GPS.txt -te /home/sagar/KAIST/test_GPS.txt -src /home/sagar/KAIST/all -s 5
Localize: scripts/vlad_place_recognition.py
Finally, this script is used to calculate the localization accuracy using the VLAD codebook. Here are it's input args -
Required
'-tr', '--train_list', type=str,required=True,help=Absolute path of the training images list with gps information. This is the same list used for building vocabulary.
'-te', '--test_list', type=str,required=True,help=Absolute path of the testing images list with gps information.
'-v', '--vlad_features', type=str,required=True,help=Absolute path of the vlad codebook file created by compute_vlad_descriptor.py
Optional
'-k', '--top_k',type=int, default=5,help="Number of nearest neighbors to retrieve"
'-d', '--acceptance_distance_thresh',type=float, default=5.0,help="Distnace threshold for localization in meters"
'-g', '--gps_type',type=str, default='meters',help="Type of gps coordinates. It can be either meters or global. global means global lat & long"
Examples
- Localization with Global gps lat & long
python vlad_place_recognition.py -tr /home/sagar/KAIST/train_GPS.txt -te /home/sagar/KAIST/test_GPS.txt -v vocabulary_f=casenet_k=64_t=0.5_alpha=0.5.vlad.npz -g global
- gps in meters
python vlad_place_recognition.py -tr /home/sagar/KAIST/train_GPS.txt -te /home/sagar/KAIST/test_GPS.txt -v vocabulary_f=casenet_k=64_t=0.5_alpha=0.5.vlad.npz
- For a specific distance threshold e.g. d=10
python vlad_place_recognition.py -tr /home/sagar/KAIST/train_GPS.txt -te /home/sagar/KAIST/test_GPS.txt -v vocabulary_f=casenet_k=64_t=0.5_alpha=0.5.vlad.npz -g global -d 10
请发表评论