Enhancing Diverse Intra-identity Representation for Visible-Infrared Person Re-Identification
Visible-Infrared person Re-Identification (VI-ReID) is a challenging task due to modality discrepancy. To reduce modality-gap, existing methods primarily focus on sample diversity, such as data augmentation or generating intermediate modality between Visible and Infrared. However, these methods do not consider the increase in intrainstance variance caused by sample diversity, and they focus on dominant features, which results in a remaining modality gap for hard samples. This limitation hinders performance improvement. We propose Intra-identity Representation Diversification (IRD) based metric learning to handle the intra-instance variance. Specifically IRD method enlarge the Intra-modality Intra-identity Representation Space (IIRS) for each modality within the same identity to learn diverse feature representation abilities. This enables the formation of a shared space capable of representing common features across hetero-modality, thereby reducing the modality gap more effectively. In addition, we introduce a HueGray (HG) data augmentation method, which increases sample diversity simply and effectively. Finally, we propose the Diversity Enhancement Network (DEN) for robustly handling intra-instance variance. The proposed method demonstrates superior performance compared to the state-of-the-art methods on the SYSU-MM01 and RegDB datasets. Notably, on the challenging SYSUMM01 dataset, our approach achieves remarkable results with a Rank-1 accuracy of 76.36% and a mean Average Precision (mAP) of 71.30%.
Keypoint Aware Robust Representation for Transformer-based Re-identification of Occluded Person
Occluded person Re-identification is a challenging task which aims to find or distinguish a specific person when the human body is occluded by obstacles, other persons or oneself. Some recent state-of-the-art works adopting a transformer and/or pose-guided methods have improved the feature representation and performances, but are still in trouble with both weak representation and heavy structure. In this paper, we suggest the novel methods of transformer-based Re-identification for the occluded person as follows. First, in data augmentation, instead of deleting an arbitrary area, only a part containing the keypoint features of a person is deleted for effective learning in occlusion. Second, we suggest a unique hierarchical patch and feature attention combining the reliability-enhanced heatmaps and the output of the transformer intermediate layer, which can more effectively pay attention to the non-occluded human region. Third, we propose an entire-partial loss function with non-similarity grouping for more robust feature representation. As a result, our suggested model not only outperforms the existing methods, but also has the smallest scale among state-of-the-art methods. We compare our approach and various existing methods for the mAP and Rank-1 performances on the Occluded-Duke, Occluded-ReID, Market-1501 and DukeMTMC datasets. Experimental results show that our proposed model outperforms the state-of-the-art methods.
Similarity-based adversarial knowledge distillation using graph convolutional neural network
This letter presents an adversarial knowledge distillation based on graph convolutional neural network. For knowledge distillation, many methods have been proposed in which the student model individually and independently imitates the output of the teacher model on the input data. Our method suggests the application of a similarity matrix to consider the relationship among output vectors, compared to the other existing approaches. The similarity matrix of the output vectors is calculated and converted into a graph structure, and a generative adversarial network using graph convolutional neural network is applied. We suggest similarity-based knowledge distillation in which a student model simultaneously imitates both of output vector and similarity matrix of the teacher model. We evaluate our method on ResNet, MobileNet and Wide ResNet using CIFAR-10 and CIFAR-100 datasets, and our results outperform results of the baseline model and other existing knowledge distillations like KLD, DML.
Boosted Knowledge Distillation Plug-in Using Controlled Knowledge Transfer
Knowledge distillation is a useful way to transfer knowledge from a teacher to a student network. The typical application is to transfer from a powerful large network to a small network, in order to meet the low-memory and/or fast execution requirements. Because conventional knowledge distillation depend only upon Kullback Leibler (KL) divergence, misled knowledge can be transferred, resulting in inaccuracy. In this paper, we suggest a controlling method to improve the efficiency of knowledge transfer. Specifically, new knowledge representation combining KL divergence and predictive accuracy of student is suggested and the knowledge are segmented to adjust the transfer rate based on the characteristics of the group. We reduce the reflection rate of losses for higher confidence of student, and increase the reflection rate of losses for lower confidence of student. By introducing sophisticated control technique, we can improve the efficiency of knowledge transfer in comparison with the existing knowledge distillation. The proposed method provides, 1) increasing the efficiency of knowledge transfer, thus improving the performance and accelerating the learning process, 2) preventing overfitting and obtaining effective regularization, 3) easily plugged in both conventional and collaborative knowledge distillation. To show the effectiveness of our approach for compression problem, we demonstrate the accuracy through performing experiments on CIFAR100, Cub-200-2011, Cars196, and Tiny ImageNet data for ResNet, MobileNet, and WRN models. We compare our results with those of existing knowledge distillation methods and obtain far better results. We also plug in ours to existing methods and to obtain much improved results than of original method. Code is available at https://github.com/ksj13/CKD.
Evolutionary Pruning of CNN Filters
Recent mobile and embedded systems re- quire e_cient visual recognition that builds on a deep learning model such as a convolutional neural network (CNN). Since features of a CNN are automatically generated for deep layers, there exists redundancy in features, which can be further compressed. Therefore, we propose evolutionary pruning algorithms to remove the irrelevant deep features in order to minimize the computational complexity and overfitting while maintaining a good quality of representation. Our proposed filter pruning via genetic algorithms (GA) helps to organize the totality of filters considering each filter's relationship and its niche among neighbors while the existing algorithmic pruning mainly focuses on weights of individual filters or activation reconstruction errors. With new GA operations included in our pruning process, we show that our GA method achieves the global optimization solution rather than a local optimization solution normally achieved by other algorithmic pruning approaches. We demonstrate a significant improvement of the filter representation by performing experiments on CIFAR-10 and smartphone metal exterior defect data. Our method is compared with results from the state-of-the-art algorithmic pruning methods in order to show its significant performance improvement by analyzing the classification performance as well as the compression rate and the oating point operations per second (FLOPS). We also show a way how to combine our computationally demanding GA method with existing fast algorithms, in order to implement our GA method to a larger network.
Genetic Programming based Gait Generation for Quadruped Robot
Planning gaits for legged robots is a challenging task that requires optimizing parameters in a highly irregular and multi-dimensional space. Several recent approaches have focused on using genetic algorithms (GAs) to generate gaits automatically and have shown significant improvement over previous gait optimization results. Most current GA-based approaches optimize only a small, pre-selected set of parameters, but it is difficult to decide which parameters should be included in the optimization to get the best results. Moreover, the number of pre-selected parameters is at least 10, so it can be relatively difficult to optimize them, given their high degree of interdependence. To overcome these problems of the typical GA-based approach, we have proposed a seemingly more efficient approach that optimizes joint trajectories instead of locus-related parameters in Cartesian space, using GP. Our GP-based method has obtained much-improved results over the GA-based approaches tested in experiments on the Sony AIBO ERS-7 in the Webots environment. The elite archive mechanism is introduced to combat the premature convergence problems in GP and has shown better results than a traditional multi-population approach.
CPG Based Gait Generation for Humanoid Robots
The research proposes generation methods for humanoid robot walking using a central pattern generator (CPG) based joint modification method to improve the adaptation ability on various slop terrains. Although CPG has basic adaptation ability for the different terrains, but it might be limited to small variations of those. In order to increase terrain adaptability we supplement the modification of joints using genetic algorithm on the output signals of the CPG oscillators. We have tested the two methods to investigate an adaptation capabilities of humanoid walking on various slope terrains with different slope angles using the humanoid robot Nao in the Webot simulation. The performances of humanoid walking on various slope terrains are analyzed.
