Edge-cloud collaborative inference can significantly reduce the delay of a deep neural network (DNN) by dividing the network between mobile edge and cloud. However, the in-layer data size of DNN is usually larger than the original data, so the communication time to send intermediate data to the cloud will also increase end-to-end latency. To cope with these challenges, this paper proposes a novel convolutional neural network structure—BBNet—that accelerates collaborative inference from two levels: (1) through channel-pruning: reducing the number of calculations and parameters of the original network; (2) through compressing the feature map at the split point to further reduce the size of the data transmitted. In addition, This paper implemented the BBNet structure based on NVIDIA Nano and the server. Compared with the original network, BBNet’s FLOPs and parameter achieve up to 5.67× and 11.57× on the compression rate, respectively. In the best case, the feature compression layer can reach a bit-compression rate of 512×. Compared with the better bandwidth conditions, BBNet has a more obvious inference delay when the network conditions are poor. For example, when the upload bandwidth is only 20 kb/s, the end-to-end latency of BBNet is increased by 38.89× compared with the cloud-only approach.
13 March 2024 in IEEE Transactions on Circuits and Systems for Video Technology
Existing research on knowledge distillation has primarily concentrated on the task of facilitating student networks in acquiring the complete knowledge imparted by teacher networks. However, recent studies have shown that good networks are not suitable for acting as teachers, and there is a positive correlation between distillation performance and teacher prediction uncertainty. To address this finding, this paper thoroughly analyzes in depth the reasons why the teacher network affects the distillation performance, gives full play to the participation of the student network in the process of knowledge distillation, and assists the teacher network in distilling the knowledge that is suitable for their learning. In light of this premise, a novel approach known as Collaborative Knowledge Distillation (CKD) is introduced, which is founded upon the concept of "Tailoring the Teaching to the Individual". Compared with Baseline, this paper’s method improves students’ accuracy by an average of 3.42% in CIFAR-100 experiments, and by an average of 1.71% compared with the classical Knowledge Distillation (KD) method. The ImageNet experiments conducted revealed a significant improvement of 2.04% in the Top-1 accuracy of the students.
Due to the large amount of video data from UAV aerial photography and the small target size from the aerial perspective, pedestrian detection in drone videos remains a challenge. To detect objects in UAV images quickly and accurately, a small-sized pedestrian detection algorithm based on the weighted fusion of static and dynamic bounding boxes is proposed. First, a weighted filtration algorithm for redundant frames was applied using the inter-frame pixel difference algorithm cascading vision and structural similarity, which solved the redundancy of the UAV video data, thereby reducing the delay. Second, the pre-training and detector learning datasets were scale matched to address the feature representation loss caused by the scale mismatch between datasets. Finally, the static bounding extracted by YOLOv4 and the motion bounding boxes extracted by LiteFlowNet were subject to the weighted fusion algorithm to enhance the semantic information and solve the problem of missing and multiple detections in UAV object detection. The experimental results showed that the small object recognition method proposed in this paper enabled reaching an mAP of 70.91% and an IoU of 57.53%, which were 3.51% and 2.05% higher than the mainstream target detection algorithm.
12 December 2023 in Computer Vision and Image Understanding