Vehicle re-identification refers to the task of retrieving all images of a particular vehicle identity in a large gallery set, composed of vehicle images taken from varying orientations, cameras, time and locations. Accurately re-identifying vehicles from images and videos, is of great interest in surveillance and intelligence applications. In contrast to vehicle recognition which aims to identify the make and model of the vehicle, vehicle re-identification is concerned with identifying specific vehicle instances. This task is extremely challenging as vehicles with different identities can be of the same make, model and color and thus it is challenging for a Deep Convolutional Neural Network (DCNN) to make accurate predictions. In this paper, we present a novel algorithm driven by adaptive attention for re-identifying vehicles from still images without using information from other sources such as time and location.
The similar task of person re-identification aims at re-identifying humans appearing in different cameras. While visual appearance models work reasonably well for person re-identification, the same techniques fail to differentiate vehicles due to the lack of highly discriminating features. Person re-identification models are not heavily reliant on facial features as they also learn discriminating features based on clothing and accessories. However, vehicle re-identification poses a new set of challenges. Different vehicle identities can have similar colors and shapes especially those coming from the same manufacturer with a particular model, trim and year. Subtle cues such as different wheel patterns and custom logos might be unavailable in the global appearance features. Therefore, it is important that vehicle re-identification model learns to focus on different parts of the vehicles while making a decision. Previous works in person re-identification such as  have used attention models with human key-points as regions of attention and have shown significant improvement in performance. Similarly, methods such as  have used vehicle key-points to learn attention maps for each of the key-points defined by . The system proposed by Wang et al. grouped key-points into four groups corresponding to front, rear, left and right.
However, not all key-points provide discriminating information and their respective contributions depend on the orientation of the vehicle. For instance, in Figure 0(a) we observe that the key-points from the front of the car incorrectly influence the attention of the model as the front of the car is not visible. Hence, paying attention to all the key-points, as suggested in , can lead to erroneous results. The proposed method tackles the problem of false attention by adaptively choosing the key-points to focus on, based on the orientation of the vehicle hence, providing complementary information to global appearance features. In this work the terms with same connotation, path, stream and branch, have been used interchangeably.
In the proposed method, the first stream is a DCNN trained to extract discriminative global appearance features for each vehicle identity. However this stream often fails to extract subtle features necessary to distinguish similar vehicles. Therefore, a second path composed of orientation conditioned key-point selection and localized feature extraction modules is used in parallel to supplement the features from the first path. By using orientation as a conditioning factor for adaptive key-point selection, the model learns to focus on the most informative parts of the vehicle. Additionally, we develop a fully convolutional two-stage key-point detection model inspired by the works of Kumaret al. and Bulat et al.
for facial key-point detection and human pose estimation respectively.
The detailed architectures of each module in the proposed method are discussed in section 3. Through extensive experimentation, we show that the proposed Adaptive Attention model for Vehicle Re-identification (AAVER) approach improves the re-identification accuracy on challenging datasets such as VeRi-776 [10, 11] and VehicleID . In addition, the proposed vehicle key-point detection model, improves the accuracy by more than over the state of the art.
2 Related Work
Similar to person re-identification, there exist subtle intra and inter-class appearance variations (i.e., orientation, color, make, model, etc.) in vehicles and therefore it is important to learn an effective and discriminative representation for accurate classification.
In this section, we briefly review recent relevant works in the field of vehicle classification and re-identification. Learning a discriminating representation, requires a large-scale annotated data for training, especially for recent deep convolutional neural network-based (DCNN) approaches. Yang et al. released a large-scale car dataset (CompCars) for fine-grained vehicle model classification which consists of 1,687 car models and 214,345 images. VehicleID dataset by Liu et al.  consists of 200,000 images of about 26,000 vehicles. In addition, Liu et al.[11, 12] published a high-quality multi-view vehicle Re-ID (VeRi-776) dataset. Yan et al. released two high-quality and well-annotated vehicle datasets, namely VD1 and VD2, with diverse annotated attributes, and containing 1,097,649 and 807,260 vehicle images captured in different cities.
Moreover, besides datasets for training, Tang et al.
claimed traditional hand-crafted features are complementary to deep features and thus fused both features to realize an improved and unified representation. Instead, Cuiet al. fused the features from various deep networks trained with different tasks and architectures. Furthermore, Liu et al.[11, 12] used multi-modal features, including visual features, license plate, camera location, and other contextual information, in a coarse to fine vehicle retrieval framework. To augment the training data for robust training,  used a generative adversarial network to synthesize vehicle images with diverse orientation and appearance variations.  learns a viewpoint-aware representation for vehicle re-identification through adverserial learning and a viewpoint-aware attention model.
Besides the global features, Liu et al.  extracted discriminative local features from a series of local regions of a vehicle by a region-aware deep model. Different from these approaches, the proposed method leverages orientation to adaptively select the regions of attentions.
Another effective strategy to learn the discriminative representation is metric learning. Zhang et al.
proposed an improved triplet loss which performs joint optimization with an auxiliary classification loss as a regularizer in order to characterize intra-sample variances. Shenet al.  also proposed to improve the matching performance by making use of spatio-temporal information; they developed a Siamese-CNN with path LSTM model which generates the corresponding candidate visual-spatio-temporal paths of an actual vehicle image by a chain Markov random field (MRF) model with a deeply learned potential function. In contrast, the proposed method uses the softmax loss function as it has shown impressive performance for the task of face verification and trains faster compared to triplet loss-based methods such as  without the hassle of sampling hard triplets.
3 Adaptive Attention Vehicle Re-identification (AAVER)
The entire pipeline of the proposed method AAVER is composed of three main modules: Global Feature Extraction, Vehicle Key-Point and Viewpoint Estimation, and Adaptive Key-Point Selection and Feature Extraction which is followed by a re-ranking based post-processing. Figure 2 shows the diagrammatic overview of our method.
In AAVER, the global feature extraction module is responsible for extracting the macroscopic features () of the vehicles. By looking at the entire vehicle, this model tries to maximally separate the identities in the feature space. However, this model may fail to take into account subtle differences between similar looking cars, most extremely the ones that are of the same make, model and color. Therefore, the features generated by this module are supplemented with features () from the localized feature extraction module. This can be achieved by the adaptive attention strategy using the proposed key-point and orientation estimation network.
In order to estimate the vehicle key-points, we draw inspiration from literature on facial key-point detection and human pose estimation. Inspired by [8, 1] we employ a two-stage model to predict the vehicle’s orientation and landmarks in a coarse to fine manner; the coarse heatmaps predicted by a deep CNN are refined using a shallower network.
Finally, we use the proposed adaptive key-point selection module to select a subset of most informative key-points and pool features from early layers of the global feature extraction module to extract localized features around the selected key-points. The features obtained from two paths of AAVER are then merged using a multilayer perceptron. The entire model can be trained end-to-end using any differentiable loss function. In our work, we use the softmax loss as proposed in . During inference, we use the features from penultimate fully connected layer as the representation of a given vehicle. Additionally, we also perform re-ranking as a post processing step.
3.1 Global Feature Extraction
For extracting the global appearance features, we employ ResNet-50 and ResNet-101  as backbone networks and also adopt them as our baseline models. We initialized the weights of these models using the weights from the models pre-trained on the Comprehensive Cars (CompCars) dataset . A
-dimensional features vector from the last convolutional layer of ResNet is then fed to a shallow multilayer perceptron with Rectified Linear Unit (RELU) non-linearity. This network is trained using thesoftmax loss function which constrains the feature vectors extracted by the network to lie on a hyper-sphere of radius . This enables the network to embed features of identical vehicles together while pushing apart the features from different vehicles. It is mathematically expressed as:
where is the feature vector corresponding to class label , is the weight and is the bias corresponding to class j, is a positive trainable scalar parameter, and is the number of classes respectively.
3.2 Vehicle Key-Point and Orientation Estimation
In this work, a two-stage model is proposed for key-point estimation. In the first stage, a VGG-16 -based fully convolutional network is employed to do a coarse estimation of the location of ( 20 key-points plus background) heatmaps of size (). This network is trained using a per-pixel multi-class cross entropy loss defined as follows:
where is the vector corresponding to pixel location and across all output channels and is the ground-truth class label for that pixel location. After training the first stage, the weights of this network are frozen for training of the subsequent stage. The left side of Figure 3 depicts the the output of the first stage for a sample vehicle image.
Although the responses of the first stage can be used for the prediction of visible key-point locations, there might be erroneous activations in the heatmaps that correspond to invisible key-points. Consequently, we use the second stage that takes in the sub-sampled version of the input image and the coarse estimates of key-points to refine the results. The refinement network follows the hourglass architecture introduced in 
which is commonly used for the refinement of heatmaps and reducing artifacts due to invisible key-points. In the second stage, coarse heatmaps estimated from the first stage, are refined through a two-stack hourglass network with skip connections. Along with refining the estimated key-points, the orientation of the vehicle is also predicted through a parallel branch composed of two fully connected layers designed, to classify the orientation into eight classes as defined in. This multitask learning helps the refinement network to make accurate predictions of the visible key-points while reducing the response of invisible key-points. Figure 3 shows the overall schematic flow of the two-stage network.
To train the heatmap refinement and orientation branches we use Mean Square Error (MSE) and cross entropy loss respectively. Equation 3 represents the loss function used for the second stage. It is worth mentioning that in the second stage we are only interested in foreground heatmaps, hence, we exclude the refinement of the background channel.
In Equation (4), , and are the stage two predicted and ground-truth heatmap for the key-point at locations and respectively. , and in Equation (5) constitute the predicted orientation vector, the corresponding ground-truth orientation and number of classes respectively. Finally in Equation (3) is a weight to balance the losses used in model optimization. In our experiments, is set to obtained after cross-validation. In the right hand side of Figure 3 which shows the the output of the second stage, it can be observed that the initial coarse estimates of key-points have been refined.
3.3 Adaptive Key-Point Selection and Feature Extraction
Subtle differences in similar looking vehicles mostly occur close to vehicle landmarks, e.g. same car make and models of same color might be distinguishable through their window stickers, rims, indicator lights on the side mirrors etc. This can be achieved by focusing the attention to parts of the image that encompasses these distinctions. To this end, regions of interest within the image are identified based on the orientation of the vehicle; after which features from the shallower layer of the global appearance model are pooled. As suggested in  these pooled features contain contextual rather than abstract information. Later, deep blocks (Res3, Res4 and Res5) of another ResNet model are used to extract supplementary features corresponding to the regions of interest.
In , vehicle’s orientation is annotated into eight different classes, i.e. rear, left ,left front, left rear, right, right front and right rear; however, there is no absolute boundary between two adjacent orientations. For instance, for the case of right and right front, the network gets confused between the two classes when trained for orientation prediction; this can be observed in Figure 4 which shows the confusion matrix for the eight class classification problem. To overcome this issue, we designed a key-point selector module that takes the predicted orientation likelihood vector and adaptively selects the key-points based on the likelihoods.
In order to achieve this, we constructed eight groups shown in Table 1
corresponding to each of the eight orientations of a vehicle and its two adjacent orientations. During inference, the likelihood of each orientation group is calculated and the one with the highest probability is picked. Also, experimentally it was observed that for each orientation group at least seven key-points are always visible. Consequently, given the orientation group with the highest probability we select the seven heatmaps shown in Table1 corresponding to the respective orientation group. These orientation groups are named based on their center orientation e.g. the group that contains left front, front and right front is named front.
|Orientation Group||Visible Key-Points|
|Front||[11, 12, 7, 8, 9, 13, 14]|
|Rear||[18, 16, 15, 19, 17, 11, 12]|
|Left||[8, 1, 11, 14, 15, 2, 17]|
|Left Front||[9, 14, 6, 8, 11, 1, 15]|
|Left Rear||[2, 17, 15, 11, 14, 19, 1]|
|Right||[7, 3, 12, 13, 16, 4, 18]|
|Right Front||[9, 13, 5, 7, 12, 3, 16]|
|Right Rear||[3, 4, 12, 16, 18, 19, 13]|
After obtaining the seven heatmaps, for each map, a Gaussian kernel with is placed in the location of the map’s peak, i.e. the key-point location. This is done in order to emphasize the importance of the surrounding areas around the key-points as they may have discriminative information.
Following the adaptive heatmap selection and dilation by the Gaussian kernel, is the localized feature extraction () by Res3, Res4 and Res5 blocks of the parallel ResNet model. The input to this sub network is the concatenation of the seven dilated heatmaps of shape and the pooled global features of shape . Finally, the localized features is concatenated with the global appearance features and passed through a multilayer perceptron followed by softmax loss function (refer to Figure 2). Considering that these features have already been normalized with
norm we use cosine similarity as the metric to calculate the similarity score between image pairs.
3.4 Post Processing Step: Re-Ranking
In general, Re-ID can be regarded as a retrieval problem. Given a probe vehicle, we want to search in the gallery for images containing the same vehicle in a cross-camera mode. After an initial ranking list is obtained, a good practice consists of adding a re-ranking step, with the expectation that the relevant images will receive higher ranks. Such re-ranking steps have been mostly studied in generic instance retrievals such as , ,  and . The main advantage of many re-ranking methods is that they can be implemented without requiring additional training samples, and also can be applied to any initial ranking list.
Significant amount of research in person re-id goes into re-ranking strategies and vehicle re-id is lacking in that aspect. Most of the state of the art methods for vehicle re-id do not perform re-ranking on their initial ranking list. We use the re-ranking strategy proposed by Zhong et al.  in our work.
Here we first present the two large-scale datasets used for the vehicle re-identification task and their evaluation protocols, after which we describe the implementation details of the proposed method.
To the best our knowledge, there are mainly two large scale vehicle datasets that are publicly available and are designed for the task of vehicle re-identification: VeRi-776 , and VehicleID .
VeRi-776 dataset consists of 49,357 images of
distinct vehicles that were captured with 20 non-overlapping cameras in variety of orientations and lighting conditions. Out of these images, 37,778 (576 identities) and 11,579 (200 identities) have been assigned to training and testing respectively. For the query set, 1,678 images have been selected from the testing set. The evaluation protocol for this dataset is as follows: for each probe image in the query set the corresponding identity and the camera ID from which the image is captured is gathered. The gallery is constructed by selecting all the images in the testing set except the ones that share the same identity and camera ID as the probe. Evaluation metrics adopted for this dataset are mean Average Precision (mAP), Cumulative Match Curve (CMC) for top 1 (CMC@1) and top 5 (CMC@5) matches.
VehicleID is another large-scale dataset used for vehicle retrieval task and is composed of images from unique vehicles. Half of the identities, i.e. , are reserved for training while the other half are dedicated for evaluation. There are test splits for gallery sizes of , , , , and . In the recent works [21, 18] the first three splits have been used frequently. The proposed evaluation protocol for each split in VehicleID dataset is to randomly select an image for each of the identities to form the gallery of respective size and use the rest of the images for query. This procedure is repeated ten times and the averaged metrics, CMC@1 and CMC@5, are reported.
4.2 Implementation Details
In our implementation, all the input images were resized to
and normalized by the ImageNet dataset
mean and standard deviation. Also, in all of our experiments we used batch training with size ofand Adam optimizer  with the learning rate of .
Initially, we fine-tuned our baseline models (see section 3.1) on VeRi-776 and VehicleID datasets separately, for epochs. then, we initialized the key-point and orientation estimation network with ImageNet pretrained weights. The first stage of this network was trained for epochs; afterwards the second stage was trained for epochs as well.
Next, we trained the orientation conditioned feature extraction branch for each of VeRi-776 and VehicleID datasets for 20 epochs. Finally, we select the network’s output of the penultimate layer as the feature vector corresponding to the input vehicle image.
5 Experimental Evaluations
We first present the evaluation results of our vehicle key-point and orientation estimation model followed by the evaluation of the proposed method AAVER on both VeRi-776 and VehicleID datasets.
5.1 Vehicle Key-Point and Orientation Estimation Evaluation
In order to evaluate the performance of the proposed two-stage key-point detection model, we use the Mean Square Error (MSE) in terms of pixels for the location of visible key-points in maps over the test set of VeRi-776 key-point dataset. Table 2 shows the MSE of our model after first and second stages. Moreover, we measured the accuracy of the model for view point classification. It can be observed that the refinement stage reduces the key-point location estimation error by compared to the first stage.
To the best of our knowledge,  is the only work on the VeRi-776 key-point and orientation estimation dataset.  used the averaged distance between estimated and ground-truth locations of all visible key-point for evaluation. If the distance is less than a threshold value ( in terms of pixels in map), the estimation is considered to be correct. We follow the same protocol to compare the precision with  and Table 2 shows the result of this comparison.
|Stage 1||Stage 2|
|Key-point location MSE (pixel)||1.95||1.56|
|Key-Point Precision Comparison|
5.2 Results on VeRi-776
Table 3 summarizes the results of global appearance model (baseline) and proposed AAVER model with adaptive attention. Note that in both ResNet-50 and ResNet-101 -based architectures, there is a significant improvement in mAP and CMC@1 scores after incorporating adaptive attention. This indicates that conditioning on the orientation of the vehicle and selecting corresponding key-points enables the network to focus more on parts that contains minute differences in similar looking cars. This claim is further studied in section 5.5. Unsurprisingly, we also observe that ResNet-101 shows better performance compared to ResNet-50 under similar settings.
Figure 5 plots the probe image and the top three returns of each baseline and the proposed model. It can be observed that AAVER significantly improves the performance over the baseline.
5.3 Results on VehicleID
Images in this dataset have less variations in viewpoint (images are either from the front or rear of the vehicles) compared to VeRi-776 dataset. For this dataset, the evaluation metrics are only CMC@1 and CMC@5 as there is only one true match in the gallery for each probe image. Table 4 presents the numerical results of baseline and the proposed models over test splits. As compared to baseline models, a significant increase in performance is observed when the features from adaptive attention-based path is fused with the global appearance features.
|Baseline Model||AAVER Model|
Figure 6 shows an examples of a query from VehicleID dataset and the top three results returned by both global and adaptive attention model.
5.4 Comparison with State of the Art Methods
In this section, we compare the AAVER model with ResNet-101 backbone against the recent state of the art methods. The results of this comparison are presented in Table 5.
|Test size = 800||Test size = 1600||Test size = 2400|
|AAVER + Re-ranking||66.35||90.17||94.34||-||-||-||-||-||-|
From Table 5, it can be observed that our proposed method is among the top performers of vehicle re-identification task and is the state of the art for most of the evaluation metrics on both VeRi-776 and VehicleID datasets. Note that in the absence of a deterministic test set for VehicleID dataset, one cannot provide the basis for a fair comparison among different methods. The reason lies in the fact that random gallery construction yields different evaluation results with relatively high variance even when averaged over ten repetitions. Finally, we have to emphasize on the necessity of using re-ranking as a post processing step whenever there are multiple instances of the probe image in the gallery. Here for the VeRi-776 dataset, re-ranking shows significant improvement and results in state of the art mAP and CMC@1 scores as compared to other methods. Note that for VehicleID dataset re-ranking is not applicable as there is only one true match in the gallery for each probe image.
5.5 Ablation Studies
We designed a set of experiments to study the impact of complementary information that the orientation conditioned branch provides. Note that in these experiments we only use the test split for the VehicleID dataset To this end, the following experiments have been conducted:
In the first experiment we examined the depth of the layer in the global branch from which the global features are pooled and then fed to the orientation conditioned branch. To investigate this we tried pooling features after Res2, Res3 and Res4 blocks of spatial size of , and . Table 6 demonstrates the result of this experiment. It can be observed that as we go from shallow to deeper layers, the features become more abstract and focusing on parts of deep feature maps do not help in providing robust representation of vehicles with minute differences.
Dataset features size mAP CMC@1 CMC@5 VeRi-776 0.612 88.97 94.70 0.608 88.50 94.58 0.597 85.88 93.03 VehicleID - 74.69 93.82 - 72.60 93.24 - 71.09 92.13 Table 6: Experimant 1: Depth of pooled global features
In our method, we proposed to use two streams for extracting global and local features from vehicle images, so we were keen to see whether a single branch can extract discriminative features that encompass global as well as local differences. To test this hypothesis, instead of pooling features from the global branch we fused the selected heatmaps into the global branch by concatenation and used the output as the representation for a vehicle image. Table 7 depicts the result of this experiment, for both VeRi and VehicleID datasets. we can infer that the re-identification performance drops significantly by relying on a single branch.
Dataset Type mAP CMC@1 CMC@5 VeRi-776 Single 0.528 80.93 90.52 Multi 0.612 88.97 94.70 VehicleID Single - 69.61 91.45 Multi - 74.69 93.82 Table 7: Experiment 2: Single versus Multi-branch feature extraction
In the final set of experiments we scrutinize the way in which the information from vehicle key-point heatmaps are incorporated in the proposed model. Our work is in some aspects similar to  which groups fixed set of key-points and combine all the corresponding heatmaps into one map by adding them together. Therefore, we conduct this experiment under the same settings as of . Table 8 shows the results of these experiments. The type ”Combined” in Table 8 refers to the method in . We can conclude that using all heatmaps combined into one group does not result in competitive results as the adaptive selection of heatmaps. This validates the hypothesis that not all the key-points contribute to a discriminative representation of the vehicle.
6 Conclusions and Future Work
In this work, we outline a robust end-to-end framework for state of the art vehicle re-identification. We present a dual path model AAVER which combines macroscopic global features with localized discriminative features to efficiently identify a probe image in a gallery of varying sizes. In addition, we establish benchmarks for key-point detection and orientation prediction on VeRi-776 dataset. Lastly, we advocate for the adoption of re-ranking when considering the performance of future vehicle re-identification methods. Adaptive key-point selection conditioned on vehicle orientation is vital for discriminating between vehicles of the same make, model and color. Evaluating on both VeRi-776 and VehicleID shows the strength of our proposed method. Lastly, we conduct an ablation study to understand the influence of the adaptive key-point selection.
In the future, we plan to extend our key-point module to align vehicle images to a canonical coordinates system before comparing a given pair of images. Similarly, we can learn a 3D representation of vehicles to be used in other tasks such as vehicular speed estimation.
A. Bulat and G. Tzimiropoulos.
Human pose estimation via convolutional part heatmap regression.
European Conference on Computer Vision, pages 717–732. Springer, 2016.
-  O. Chum, J. Philbin, J. Sivic, M. Isard, and A. Zisserman. Total recall: Automatic query expansion with a generative feature model for object retrieval. In 2007 IEEE 11th International Conference on Computer Vision, pages 1–8. IEEE, 2007.
-  C. Cui, N. Sang, C. Gao, and L. Zou. Vehicle re-identification by fusing multiple deep neural networks. In International Conference on Image Processing Theory, Tools and Applications (IPTA), pages 1–6, 2017.
-  J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. 2009.
-  K. He, X. Zhang, S. Ren, and J. Sun. Deep residual learning for image recognition. CoRR, abs/1512.03385, 2015.
H. Jegou, H. Harzallah, and C. Schmid.
A contextual dissimilarity measure for accurate and efficient image
2007 IEEE Conference on Computer Vision and Pattern Recognition, pages 1–8. IEEE, 2007.
-  D. P. Kingma and J. Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980, 2014.
-  A. Kumar and R. Chellappa. Disentangling 3d pose in a dendritic cnn for unconstrained 2d face alignment. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018.
-  H. Liu, Y. Tian, Y. Wang, L. Pang, and T. Huang. Deep relative distance learning: Tell the difference between similar vehicles. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 2167–2175, 2016.
-  X. Liu, W. Liu, H. Ma, and H. Fu. Large-scale vehicle re-identification in urban surveillance videos. In IEEE International Conference on Multimedia and Expo (ICME), pages 1–6, 2016.
-  X. Liu, W. Liu, T. Mei, and H. Ma. A deep learning-based approach to progressive vehicle re-identification for urban surveillance. In European Conference on Computer Vision, pages 869–884. Springer, 2016.
-  X. Liu, W. Liu, T. Mei, and H. Ma. Provid: Progressive and multimodal vehicle reidentification for large-scale urban surveillance. IEEE Transactions on Multimedia, 20(3):645–658, 2018.
-  X. Liu, S. Zhang, Q. Huang, and W. Gao. Ram: a region-aware deep model for vehicle re-identification. In 2018 IEEE International Conference on Multimedia and Expo (ICME), pages 1–6. IEEE, 2018.
-  A. Newell, K. Yang, and J. Deng. Stacked hourglass networks for human pose estimation. In European Conference on Computer Vision, pages 483–499. Springer, 2016.
-  A. Paszke, S. Gross, S. Chintala, G. Chanan, E. Yang, Z. DeVito, Z. Lin, A. Desmaison, L. Antiga, and A. Lerer. Automatic differentiation in pytorch. In NIPS-W, 2017.
-  D. Qin, S. Gammeter, L. Bossard, T. Quack, and L. Van Gool. Hello neighbor: Accurate object retrieval with k-reciprocal nearest neighbors. In CVPR 2011, pages 777–784. IEEE, 2011.
-  R. Ranjan, C. D. Castillo, and R. Chellappa. L2-constrained softmax loss for discriminative face verification. arXiv preprint arXiv:1703.09507, 2017.
-  Y. Shen, T. Xiao, H. Li, S. Yi, and X. Wang. Learning deep neural networks for vehicle re-id with visual-spatio-temporal path proposals. In IEEE International Conference on Computer Vision (ICCV), pages 1918–1927, 2017.
-  K. Simonyan and A. Zisserman. Very deep convolutional networks for large-scale image recognition. arXiv preprint arXiv:1409.1556, 2014.
-  Y. Tang, D. Wu, Z. Jin, W. Zou, and X. Li. Multi-modal metric learning for vehicle re-identification in traffic surveillance environment. In IEEE International Conference on Image Processing (ICIP), pages 2254–2258, 2017.
-  Z. Wang, L. Tang, X. Liu, Z. Yao, S. Yi, J. Shao, J. Yan, S. Wang, H. Li, and X. Wang. Orientation invariant feature embedding and spatial temporal regularization for vehicle re-identification. In IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 379–387, 2017.
F. Wu, S. Yan, J. S. Smith, and B. Zhang.
Joint semi-supervised learning and re-ranking for vehicle re-identification.In IEEE Conference on Pattern Recognition (ICPR), 2018.
-  J. Xu, R. Zhao, F. Zhu, H. Wang, and W. Ouyang. Attention-aware compositional network for person re-identification. In The IEEE Conference on Computer Vision and Pattern Recognition (CVPR), June 2018.
-  K. Yan, Y. Tian, Y. Wang, W. Zeng, and T. Huang. Exploiting multi-grain ranking constraints for precisely searching visually-similar vehicles. In IEEE International Conference on Computer Vision (ICCV), 2017.
-  L. Yang, P. Luo, C. Change Loy, and X. Tang. A large-scale car dataset for fine-grained categorization and verification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 3973–3981, 2015.
-  M. D. Zeiler and R. Fergus. Visualizing and understanding convolutional networks. CoRR, abs/1311.2901, 2013.
-  Y. Zhang, D. Liu, and Z.-J. Zha. Improving triplet-wise training of convolutional neural network for vehicle re-identification. In IEEE International Conference on Multimedia and Expo (ICME), pages 1386–1391, 2017.
-  Z. Zhong, L. Zheng, D. Cao, and S. Li. Re-ranking person re-identification with k-reciprocal encoding. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, pages 1318–1327, 2017.
-  Y. Zhou and L. Shao. Viewpoint-aware attentive multi-view inference for vehicle re-identification. In Proc. IEEE Conf. Comp. Vis. Patt. Recogn, volume 2, 2018.