As a cost-effective and robust technology, automotiveradar has seen steady improvement during the last years.Radio frequency (RF) images, serving as a radar data formatwith rich semantic information, have attracted considerableinterest in radar object detection. Previous RF-based modelsheavily rely on convolutional neural networks, leading to thehigh computational cost. To solve this problem, we propose amodel called Mask-RadarNet to fully utilize the hierarchicalsemantic features from the RF image sequences. Mask-RadarNetexploits the combination of interleaved convolution and attentionoperations in the encoder. In addition, patch shift is introducedto Mask-RadarNet for efficient spatial-temporal feature learning.By shifting part of patches with a specific mosaic pattern inthe temporal dimension, Mask-RadarNet achieves competitiveperformance while reducing the computational burden of thespatial-temporal modeling. In order to capture the spatial-temporal semantic contextual information, we design the classmasking attention module (CMAM) in our encoder. Moreover,a lightweight auxiliary decoder is added to our model toaggregate prior maps generated from the CMAM. Experimentson the CRUW dataset demonstrate that the proposed Mask-RadarNet achieves state-of-the-art performance with relativelylower computational complexity and fewer parameters.

4D millimeter-wave radar has demonstrated significant potential for 3D object detection in autonomous drivingdue to its cost-effectiveness and robustness. However, the inherent sparsity of radar data poses a significantchallenge to achieving accurate 3D object detection, as it limits the amount of meaningful information availablefor feature learning. The key to addressing the performance degradation caused by sparsity lies in expandingthe receptive field and enriching feature representations. To this end, we propose a multi-scale, multi-layer,multi-frame network with a large receptive field, named Radar M2- Net. First, we design a multi-scale voxelfeature encoding (MSVFE) module and a multi-layer attention (MLA) module, both of which significantlyexpand the receptive field and enrich features, effectively addressing the issue of sparsity. Then, a multi-frame fusion module is developed to further enhance features by utilizing the accumulation of temporalinformation. Simultaneously, we design a novel sparse dual-head within the sparse framework to address thedecline in detection accuracy for large object categories caused by radar sparsity. Extensive experiments onthe View-of-Delft and TJ4DRadSet datasets have confirmed the advancement and effectiveness of our network.Specifically, our method achieves state-of-the-art mean average precision (mAP) performance on both datasets,even outperforming some multimodal approaches in certain metrics.