When training DETR-like models What matters... (1) classification branch initialization.initialize the bias with prob. 0.1 when ti comes to focal loss (2) grid mask augmentationaugment input images using random grid masking 더보기 Useful CursorAI scripts ## Make standalone classes or functions from existing libaries (mmdet, mmdet3d_plugin, etc) Given @maptr_tiny_r50_24e.py (129-150) and the codes in this project, make a class that runs standalone without relying on mmdetection3d and mmdet libraries and implements @decoder.py (7-83) . After making the class, make a validation code that the original implementation and your standalone implementatio.. 더보기 nuScenes ego-centric visualization via opencv # For visualization ------------------ # NuScenes convention: x = forward, y = left, z = up # BEV image layout: top = forward (+x), left = left (+y) x_range = (self.pc_range[0], self.pc_range[3]) y_range = (self.pc_range[1], self.pc_range[4]) map_size = 512 axis_range_y = y_range[1] - y_range[0] axis_range_x = x_range[1] - x_range[0] scal.. 더보기 이전 1 2 3 4 ··· 72 다음