Computer Vision News - August 2022

6 AI Research The new dataset is manually labelled and uses a high-resolution lidar that can dynamically increase point cloud resolution around key objects and a camera Intel RealSense D435i, mounted directly below the lidar. The cars are labelled using the tool SUSTechPOINTS and only if they are visible in the image FOV. The experiments include: • Domain adaptation between nuScenes -> KITTI: addressing difference in lidar ring numbers • Domain adaptation between Waymo -> KITTI: addressing the use of multiple concatenated point clouds to single point cloud • Domain adaptation between nuScenes/Waymo -> Baraja: ring-based to a uniform, interleaved scan pattern On the above three tasks, the authors compare the following methods: 1) source-only (no domain adaptation), 2) ST3D (state-of-the-art method on 3D object detection using self- training, 3) SEE with segmentation algorithm for object isolation (step 1 in Figure 1), 4) SEE-Ideal with ground truth annotations to isolate the target domain objects, 5) Oracle, the fully supervised detector trained on the target domain. For validating the methods, the 3D detectors PointVoxel-RCNN and SECOND-IoU are used. Figure 4: Example of Baraja Spectrum-Scan™ dataset Figure 3: Difference between public datasets (a), KITTI ring separation (b)