Cenet: Toward Concise and Efficient Lidar Semantic Segmentation for Autonomous Driving

Rethinking Range View Representation for LiDAR Segmentation

“…The advantages of RangeFormer and STR are demonstrated from aspects of accuracy and efficiency on prevailing benchmarks. Concretely, we achieve 73.3% mIoU and 64.2% PQ on SemanticKITTI [4], surpassing prior range view methods [12,82] by significant margins and also better than SoTA fusion-based methods [28,74,76]. We also establish superiority on nuScenes [19] (sparser point clouds) and ScribbleKITTI [65] (weak supervisions) which validates our scalability.…”

“…Ef-ficientLPS [55] proposed a proximity convolution module to leverage neighborhood points in the range image. FID-Net [82] and CENet [12] switch the encoders to ResNet and replace the decoder with simple interpolations. In contrast to using FCNs, we build RangeFormer upon self-attentions and demonstrate potential and advantages for long-range dependency modeling in range view learning.…”

“…[45,64,71]. Unfortunately, the segmentation accuracy of current projection-based methods [12,80,82] is still far behind the trend [72,74,76]. The challenge of learning from projected LiDAR scans comes from the potential detrimental factors of the LiDAR data representation [45].…”

“…While these problems are ubiquitous in range view learning, previous works hardly consider tackling them. Stemming from the image segmentation community [79], prior arts widely adopt the fully-convolutional networks (FCNs) [7,43] for range view LiDAR segmentation [12,33,45,82]. The limited receptive fields of FCNs cannot directly model long-term dependencies and are thus less effective in handling the mentioned impediments.…”

“…Inspired by the success of Vision Transformer (ViT) and its followups [17,41,57,67,70], we design a new framework dubbed RangeFormer to better handle the learning and processing of LiDAR point clouds from the range view. We formulate the segmentation of range view grids as a seq2seq problem and adopt the standard self-attention modules [66] to capture the rich contextual information in a "global" manner, which is often omitted in FCNs [1,12,45]. The hierarchical features extracted with such global awareness are then fed into multi-layer perceptions (MLPs) for decoding.…”