RADDet: Range-Azimuth-Doppler based Radar Object Detection for Dynamic Road Users

Raw High-Definition Radar for Multi-Task Learning

“…While they provide accurate range and velocity, they suffer from a low azimuth resolution leading to ambiguity in separating close objects. Recent datasets include processed radar representations such as the entire Range-Azimuth-Doppler (RAD) tensor [31], [43] or single views of this tensor -either Range-Azimuth (RA) [1], [38], [17], [41], [27] or Range-Doppler (RD) [27]. These representations require large bandwidth to be transmitted as well as large memory storage.…”

“…Specific architectures have been designed to ingest aggregated views of the RAD tensor to detect objects in the RA view [23], [11]. The entire tensor has also been considered, either for object detection in both RA and RD views [43] or for object localisation in the camera image [32].…”

“…Owing to the promising capabilities of HD radars, our work [24] 2019 Small HD CL 3D Boxes RadarRobotCar [1] 2020 Large S CLO CARRADA [31] 2020 Small LD C Segmentation RADIATE [38] 2020 Medium S CLO 2D Boxes MulRan [17] 2020 Medium S CLO Zendar [27] 2020 Small HD CL 2D Boxes CRUW [41] 2021 Medium LD C Point Location RadarScenes [36] 2021 Large HD CO Point-wise RADDet [43] 2021…”

GenRadar: Self-supervised Probabilistic Camera Synthesis based on Radar Frequencies

“…Some sets expose low sample rates and imbalanced object class distributions whereas other records do not include Doppler information [15], [16], depriving themselves of radars unique feature altogether. Only recently, first radar datasets featuring annotations on the frequency level were made publicly available [17], [18], [19] but it remains to be seen if these are going to have a similar impact on the community and will incite comparable research efforts as the famous KITTI [20] and Cityscapes [21] benchmarks did for vision-based scene-understanding. Both aspects, the tedious annotation of radar data followed by an inevitable inflow of misinformation and the preferable elimination of equivocation call for completely different approaches, establishing the subfield of self-supervised learning.…”

“…As no resampling is applied to correct for vastly misclassified first camera tokens, the importance of drawing high-quality samples cannot be overstated. Introducing a temperature parameter into the softmax normalization of the raw transformer logits akin to equation (18) and truncating the tails of the PMF by topk selection, enables so-called nucleus sampling [85] of the camera tokens. More precisely, shrinking the sample space to only a few categories K K comprising the bulk of the probability mass increases both the sample quality and reliability by preventing low-probability outcomes.…”