Multi-view deep neural network is perhaps the most successful approach in 3D shape classification. However, the fusion of multi-view features based on max or average pooling lacks a view selection mechanism, limiting its application in, e.g., multi-view active object recognition by a robot. This paper presents VERAM, a recurrent attention model capable of actively selecting a sequence of views for highly accurate 3D shape classification. VERAM addresses an important issue commonly found in existing attention-based models, i.e., the unbalanced training of the subnetworks corresponding to next view estimation and shape classification. The classification subnetwork is easily overfitted while the view estimation one is usually poorly trained, leading to a suboptimal classification performance. This is surmounted by three essential view-enhancement strategies: 1) enhancing the information flow of gradient backpropagation for the view estimation subnetwork, 2) devising a highly informative reward function for the reinforcement training of view estimation and 3) formulating a novel loss function that explicitly circumvents view duplication. Taking grayscale image as input and AlexNet as CNN architecture, VERAM with 9 views achieves instance-level and class-level accuracy of 95.5% and 95.3% on ModelNet10, 93.7% and 92.1% on ModelNet40, both are the state-of-the-art performance under the same number of views.

Wearing a helmet is one of the effective measures to protect workers' safety. To address the challenges of severe occlusion, multi-scale, and small target issues in helmet detection, this paper proposes a helmet detection algorithm based on deformable attention transformers. The main contributions of this paper are as follows. A compact end-to-end network architecture for safety helmet detection based on transformers is proposed. It cancels the computationally intensive transformer encoder module in the existing detection transformer (DETR) and uses the transformer decoder module directly on the output of feature extraction for query decoding, which effectively improves the efficiency of helmet detection. A novel feature extraction network named Swin transformer with deformable attention module (DSwin transformer) is proposed. By sparse cross-window attention, it enhances the contextual awareness of multi-scale features extracted by Swin transformer, and keeps high computational efficiency simultaneously. The proposed method generates the query reference points and query embeddings based on the joint prediction probabilities, and selects an appropriate number of decoding feature maps and sparse sampling points for query decoding, which further enhance the inference capability and processing speed. On the benchmark safety-helmet-wearing-dataset (SHWD), the proposed method achieves the average detection accuracy mAP@0.5 of 95.4% with 133.35G floating-point operations per second (FLOPs) and 20 frames per second (FPS), the state-of-the-art method for safety helmet detection.

As privacy regulations grant data owners the “right to be forgotten”, it is becoming increasingly critical for federated learning (FL) to enable the removal of individual data contributions to the global model. This process, referred to as unlearning, is essential to protect user privacy in distributed machine learning scenarios. However, the development of unlearning in the FL context still faces two main challenges: 1) memory residual, and 2) excessive resource overheads. In this paper, we propose a novel federated unlearning method via adjusting model parameters to overcome these challenges. Specifically, we design a discriminant matrix to accurately identify the parameters associated with the data to be unlearned. Furthermore, we present a scaling approach to adjust these parameters, preventing significant negative impact on the global model during unlearning. Finally, the experimental results demonstrate that our proposed method achieves comparable unlearning effectiveness to the gold standard of retraining from scratch, and significantly improving the efficiency of the unlearning. It improves efficiency by up to 9× relative to full retraining, with only about a 2% degradation in unlearning effectiveness.

As an extension of machine unlearning in distributed scenarios, federated unlearning gains significant attention. However, federated unlearning remains challenging, as many studies require additional resources, such as auxiliary dataset or storage, to achieve high-quality models. These requirements incur extra costs and are often difficult to satisfy in practical applications. To address these issues, we propose a flexible client-level federated unlearning algorithm with prototypes, called FedUP. Specifically, our algorithm consists of two components: prototype-based unlearning and model recovering. First, we design a prototype-based unlearning strategy that uses prototypes of the erased client to guide the unlearning process, and maximizes the prototype loss between the remaining and erased clients to unlearn the information. It does not rely on historical storage updates or additional standard datasets, making the unlearning process more streamlined. To mitigate performance degradation from the unlearning process, we develop a brief model recovering approach guided by global prototypes to swiftly and efficiently restore models' accuracy on the remaining datasets. Unlike other unlearning algorithms, our approach exchanges prototypes instead of model parameters, significantly reducing communication overhead. Finally, we empirically evaluate the proposed algorithm from multiple perspectives on two datasets, demonstrating that our algorithm can achieve high-quality unlearned models with minimal communication cost.

Visual browse and exploration in motion capture data take resource acquisition as a human–computer interaction problem, and it is an essential approach for target motion search. This paper presents a progressive schema which starts from pose browse, then locates the interesting region and then switches to online relevant motion exploration. It mainly addresses three core issues. First, to alleviate the contradiction between the limited visual space and ever-increasing size of real-world database, it applies affinity propagation to numerical similarity measure of pose to perform data abstraction and obtains representative poses of clusters. Second, to construct a meaningful neighborhood for user browsing, it further merges logical similarity measures of pose with the weight quartets and casts the isolated representative poses into a structure of phylogenetic tree. Third, to support online motion exploration including motion ranking and clustering, a biLSTM-based auto-encoder is proposed to encode the high-dimensional pose context into compact latent space. Experimental results on CMU’s motion capture data verify the effectiveness of the proposed method.

Multi-view deep neural network is perhaps the most successful approach in 3D shape classification. However, the fusion of multi-view features based on max or average pooling lacks a view selection mechanism, limiting its application in, e.g., multi-view active object recognition by a robot. This paper presents VERAM, a recurrent attention model capable of actively selecting a sequence of views for highly accurate 3D shape classification. VERAM addresses an important issue commonly found in existing attention-based models, i.e., the unbalanced training of the subnetworks corresponding to next view estimation and shape classification. The classification subnetwork is easily overfitted while the view estimation one is usually poorly trained, leading to a suboptimal classification performance. This is surmounted by three essential view-enhancement strategies: 1) enhancing the information flow of gradient backpropagation for the view estimation subnetwork, 2) devising a highly informative reward function for the reinforcement training of view estimation and 3) formulating a novel loss function that explicitly circumvents view duplication. Taking grayscale image as input and AlexNet as CNN architecture, VERAM with 9 views achieves instance-level and class-level accuracy of 95:5% and 95:3% on ModelNet10, 93:7% and 92:1% on ModelNet40, both are the state-of-the-art performance under the same number of views.