Dite-HRNet: Dynamic Lightweight High-Resolution Network for Human Pose Estimation

Abstract

Accurate and real-time LiDAR semantic segmentation is necessary for advanced autonomous driving systems. To guarantee a fast inference speed, previous methods utilize the highly optimized 2D convolutions to extract features on the range view (RV), which is the most compact representation of the LiDAR point clouds. However, these methods often suffer from lower accuracy for two reasons: 1) the information loss during the projection from 3D points to the RV, 2) the semantic ambiguity when 3D points labels are assigned according to the RV predictions. In this work, we introduce an end-to-end point-range fusion network (PRNet) that extracts semantic features mainly on the RV and iteratively fuses the RV features back to the 3D points for the final prediction. Besides, a novel range view projection (RVP) operation is designed to alleviate the information loss during the projection to the RV, and a point-range convolution (PRConv) is proposed to automatically mitigate the semantic ambiguity during transmitting features from the RV back to 3D points. Experiments on the SemanticKITTI and nuScenes benchmarks demonstrate that the PRNet pushes the range-based methods to a new state-of-the-art, and achieves a better speed-accuracy trade-off.

Lightweight high-resolution network based on adaptive cross-dimensional weighting for human pose estimation

“…Lite-HRNet 4 introduces a conditional channel weighting module consisting of channel attention mechanisms to substitute the computationally expensive pointwise convolutions in the shuffle block. Dite-HRNet 5 imparts dynamic attributes to the convolutional kernels, improving network computational efficiency and enhancing long-range spatial dependencies. However, the attention module adopted by the dynamic split convolution in Dite-HRNet closely resembles its proposed global context modeling module.…”

“…However, its heightened complexity presents challenges for implementation on resource-constrained devices. Recent research endeavors 4 – 7 have been devoted to enhancing the efficiency of networks. Small HRNet 1 reduces the depth and width of HRNet 3 .…”

“…Nevertheless, this fixed network structure, independent of the input, somewhat constrained the model’s performance. Dite-HRNet 5 introduces dynamic split convolution embedded within dynamic lightweight blocks, resulting in a further improvement in network computational efficiency. Our research centers on the design of input-dependent adaptive modules, recognizing that specific operations may yield differing effects at various positions within the network and under varying input sizes 8…”

Lightweight Joint Loss 2D Pose Estimation Network Based onCM-RTMPose

“…Top-down pose estimation algorithms [15,16] belong to two-stage networks, which locate the specific position and size of instances through object detection, and then perform fine-grained estimation using single-person pose estimation methods to obtain the keypoints of the entire image instance. Mask RCNN [17] extracts image features, generates a series of candidate regions through Region Proposal Network (RPN), and then performs classification and keypoint prediction for each candidate region.…”

“…LiteHRNet [20] applies the shuffle block in ShuffleNet to HRNet to produce stronger performance in lightweight networks, while improving model inference speed. Dite-HRNet [15] proposes dynamic multi-scale context blocks and dynamic global context blocks to reduce model parameters while ensuring model accuracy. UULPN [21] designs a lightweight bottleneck module with a reparameterization structure to improve model accuracy while maintaining the same computational cost.…”