Person Re-Identification by Symmetry-Driven Accumulation of Local Features, CVPR 2010
We proposed an appearance-based method for person re-identification. It consists in the extraction of features that model three complementary aspects of the human appearance: the overall chromatic content, the spatial arrangement of colors into stable regions, and the presence of recurrent local motifs with high entropy. All this information is derived from different body parts, and weighted opportunely by exploiting symmetry and asymmetry perceptual principles. In this way, robustness against very low resolution, occlusions and pose, viewpoint and illumination changes is achieved. The approach applies to situations where the number of candidates varies continuously, considering single images or bunch of frames for each individual.
Download SDALF (v0.3, 16/02/2012)
Download Partitions
Download CMC curves
- For the instructions, have a look at the README.txt file
Datasets:
[1] VIPeR http://vision.soe.ucsc.edu/?q=node/178 provided by: D. Gray, S. Brennan, and H. Tao. Evaluating appearance models for recongnition, reacquisition and tracking. In Proceedings of the IEEE International Workshop on Performance Evaluation of Tracking and Surveillance (PETS), 2007.
[2] iLIDS provided by: W. Zheng, S. Gong, and T. Xiang. Associating groups of people. In BMVC, 2009. (you will need the iLIDS MCTS license)
[3] ETHZ http://www.umiacs.umd.edu/~schwartz/datasets.html provided by: A. Ess, B-Leibe, and L. V. Gool. Depth and appearance for mobile scene analysis. In IEEE International Conference on Computer Vision, 2007.
Bibtex:
@inproceedings{Farenzena:2010CVPR,
author = {Farenzena, Michela and Bazzani, Loris and Perina, Alessandro and Murino, Vittorio and Cristani, Marco},
title = {Person Re-Identification by Symmetry-Driven Accumulation of Local Features},
booktitle = {Proceedings of the 2010 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR 2010)},
year = {2010},
publisher = {IEEE Computer Society},
address = {San Francisco, CA, USA},
}
Thanks for your work!But I have a problem : for ETHZ datasets image is named as “frame0713Person07.png”, and you named image as “0007025.png “.How did you generate “0007025.png” from “frame0713Person07.png”. In my opinion 0007 represent 0007th person, and 025 repersent 025th image of 0007th person,but 025th image is not equal to 025th frame of 0007th person.
Hi Paul, I renamed all the filenames for that dataset, in order to have a coherent filename format for all the datasets.
The standard I used is: (XX,YY)
where XX are 4 digits for the person ID and YY are 3 digits of the ID for the images
For example, 0007118.png corresponds to the 118th image of the 7th pedestrian.
I will add more instructions in the README.txt file of the package.
Thank you for your useful comment.
Hi Loris,
I’m running the code on the VIPeR dataset and getting lower CMC results than in the paper (see hoan.gs/files/temp/viper-cmc.png). Does the code already take into account making random sets of 316 people (as in the paper), or are the lower results because it is doing a single run on the entire 632-person dataset (or some other reason)?
Thanks for the work!
Hi Hoang, the code that I have uploaded here was not the same of our CVPR paper. In fact, the cross-validation is done on the entire dataset (632 people) for VIPeR. The main reason is because I have tried to make the cross-validation “dataset-independent”. However, in order to compare SDALF with the other methods, we have to do different cross-validations (see our paper), one for each dataset.
I have just replaced the code with the original version, because now I do not have time to correct the bugs. I have also provided the pre-computed features for VIPeR in the MAT/ folder. I hope this will help you.
Thank you for giving me feedbacks about the code.
Hi Loris,
Thank you for the updated code and helpful response!
From your experience, which of the 3 features (MSCR, WHSV, RHSP) do you feel contribute most to performance (individually or as a complementary feature)?
Thanks again,
Hoang
The most of the work is done by the weighted HSV. Still, using the symmetry-based body segmentation and the weighting technique is very important to improve the accuracy with respect to the usual HSV histogram. After that, we have the MSCR descriptor. The minor contribution is given by RHSP, because it is discriminative when people dress clothes with patterns.
In the paper, RHSP extracts the HSV histograms,but in the code extracts the LBP feature.
I’m eager to know what difference it makes. whether have you extracted the HSV histograms in your previous work. If you had, what’was the result?
hope for your response!
Best wishes to you
The current version of the code is not the original one, because we have modified it in order to improve it and to make it more clear to understand. We replaced HSV histograms with LBP histograms in order to capture texture information from the data. The results in terms of CMC are very similar, but intuitively the color is already covered by the other feature (wHSV). With LBP we would like to have additional information.
I hope this replies to your question. If not, please send me an email.
Thanks for your response.
I’m very interesting in your research. As you have said, wHSV is the most efficient feature for the match, if you tell me the CMC curve obtained only used wHSV(or MSCR or RHSP), will help me a lot.
hope for your response.
Hi Loris,
could you explain me the relation between the names of images specified in VIPeR_sets.txt (for example cam_a/001_45.bmp cam_b/001_90.bmp ) and the names as 0001001.bmp, 0001002.bmp ecc?
Thanks.
Paul
We renamed the dataset as (XX,YY), where XX are 4 digits for the person ID and YY are 3 digits of the ID for the images. Since in VIPeR there are two images for each person, we will have that YY=001 for cam_a and YY=002 for cam_b.
In addition, VIPeR IDs start from 0 while ours dataset from 1. For example, cam_a/005_45.bmp => 0006001.bmp (note 005 and 0006).