Recently, Pensees's Vehicle ReID technology refreshed a world record on the vehicle re-identification dataset VERI-Wild in the Wild and broke the best results of the VCIP 2019 Vehicle Re-Identification Challenge. Based on an in-house developed global and local features deep fusion network, Pensees has achieved a significant increase of key indicators of the vehicle re-recognition algorithm, the Mean Index Precision (mAP) and Rank-1 Accuracy. The mAP on the VERI-Wild dataset reached 85.35%.
Vehicle ReID has a broad application prospect and plays a vital role in intelligent transport and public safety & security and has been the focus of research in the field of computer vision. Pensees continues to deepen the development and innovation of scenario-oriented AI technology and has made breakthroughs in vehicle re-identification in non-restricted scenarios, demonstrating the company's computer vision technology strength in implementation.
01.
Challenges of vehicle ReID and the launch of the VERI-Wild dataset
Vehicle ReID, also known as vehicle re-identification, aims to find the same vehicle in different surveillance scenes. Recently, with the development of deep learning technology, the efficiency of vehicle re-identification algorithms has improved significantly. However, the limitations of existing datasets have oversimplified the practical challenges faced by vehicle re-identification, making the ReID model developed and evaluated based on most existing datasets potentially problematic in real-world scenarios. Vehicle ReID in real surveillance scenarios still faces challenges such as height viewpoint variations, extreme illumination conditions, background clutter, and different camera sources. The launch of the VERI-Wild dataset for vehicle re-identification in the Wild is dedicated to solving these problems.
Comparison of samples on VERI-Wild and Vehicle ID, VeRI-776 datasets
VERI-Wild is a vehicle re-identification dataset proposed during CVPR 2019 conference. The dataset consists of 400,000 images across 40,000 vehicle IDs and also extra information such as vehicle brands, colors and types that can be used to enhance the performance of the ReID framework or as an independent task. The VERI-Wild dataset is designed to address the limitations of existing datasets such as a limited number of vehicle identities and images, limited number of cameras and coverage areas, limited camera height viewpoint, and no significant changes in illumination and weather conditions. It's the most challenging vehicle re-identification dataset.
02.
Pensees proposed a global and local features deep fusion network for vehicle ReID
At the Grand Challenges on Vehicle ReIdentification held during the 2019 IEEE International Conference on Visual Communications and Image Processing (VCIP), Pensees proposed a global and local features deep fusion network for vehicle re-identification. Jane Shen, Chief Scientist of Pensees Technology and Dean of Singapore Research Institute, was invited to attend the conference and gave a report on "Global and Local Deep Feature Representation Fusion for Vehicle Re-Identification".
Jane Shen, Chief Scientist of Pensees Technology and Dean of Singapore Research Institute, was invited to attend VCIP 2019
Many vehicles with different logos had extremely similar appearances. Through various methods, Pensees's algorithm team made use of various parts of the vehicle to select model-based features to perform model predictions. In this way, the model could better know the unique features of components.
The global and local features deep fusion network proposed by Pensees
Recently, the vehicle ReID algorithm team of Pensees's Singapore Research Institute considered the use of feature vectors (without classification layer) in the ReID task to calculate the distance matrix, and then compares the similarity between the two images. Classification loss itself is not enough to train a good model. Then, the team applied deep metric learning (DML) onto the latest model as to force the distance between intra-class triplets less than the distance between inter-class ones by at least a certain margin, thereby improving the performance of the model.
After testing, the vehicle ReID algorithm model proposed by Pensees was significantly better than the baseline model in three test sets of different sizes of VERI-Wild. The Mean Index Precision (mAP) and the Rank-1 accuracy were significantly improved and refreshed world record.
[1] VERI-Wild: A Large Dataset and a New Method for Vehicle Re-Identification in the Wild
Evaluation results on vehicle re-identification dataset VERI-WILD
Similarly, the performance of this model was also better than the team from the Institute of Automation of the Chinese Academy of Sciences which ranked the first on the VCIP 2019 Grand Challenges on Vehicle ReIdentification.
Comparison of test results of VCIP 2019 Vehicle Re-Identification Challenge
03.
ReID algorithm is continuously optimized to deepen the development and innovation of scenario-oriented AI technology
ReID is the most direct method for resolving target rematching after losing it in cross-camera tracking. It is a very effective feature for multi-target and single-target tracking in single camera. Pensees has accumulated rich algorithm experience in target ReID, achieved many world-class results, and gradually realized the application of ReID technology in the smart city construction scenario.
This year, Pensees has made breakthroughs in Person ReID and Video-based Person ReID. In July, Pensees was tested on three mainstream datasets, Market1501, DukeMTMC-reID, and CUHK03, and the key indicator Rank-1 Accuracy achieved the best in the industry. In August, Pensees also set a new world record on three major datasets PRID-2011, iLIDS-VID, and MARS for video-based person ReID, and achieved a significant improvement of key indicators of algorithms.
The achievements made by Pensees on the dataset for Vehicle ReID in the wild also confirmed the company's insistence on the development of scenario-oriented AI technology. Vehicle ReID in the wild is closer to the actual application scenario. With its self-developed algorithm model, Pensees effectively improved the algorithm's performance under realistic challenges such as height viewpoint variations, extreme illumination conditions, background clutter, and different camera sources. Next, Pensees will gradually realize the implementation of the algorithm in the fields of safe cities and intelligent transportation.
The AI industry enters the stage of industrialization dominated by commercial implementations. As an AI company which focuses on computer vision and IoT technologies and provides "people-centric" comprehensive application solutions, it will continue to deepen the scenario-oriented development and innovation of AI technology, deepening into the scenarios according to user needs, discover pain points of the industry and continue to polish algorithms and products, apply technologies to business scenarios, and promote the production and commercialization of AI technology.
<The End>
Comments