Chase Wu

huge image测试问题反馈

6

1. 请问`BT_nms`是什么，我更换成`obb_nms`（根据代码应该是有gpu加速）后，nms速度快了不少 2. `mmdet/models/detectors/obb/obb_base.py`中`img = img.copy()`一行（第24行）消耗了很多运行时间，由于我的图片较大（15000*15000），程序会在这里卡住 3. `BboxToolkit/BboxToolkit/visualization/show.py`由于采用pyplot，同样有上述卡住的问题，暂时不知道怎么解决

This pytorch version: `return logits` Deepmind's version: ``` averaged_logits = tf.reduce_mean(logits, axis=1) end_points[end_point] = averaged_logits if self._final_endpoint == end_point: return averaged_logits, end_points end_point = 'Predictions' predictions = tf.nn.softmax(averaged_logits) end_points[end_point] =...

https://github.com/OpenDriveLab/UniAD/blob/2f38ff1357d3956af11c5609d5275db56c559c20/projects/mmdet3d_plugin/uniad/detectors/uniad_e2e.py#L263-L283 On line 267, update `self.prev_frame_info['scene_token']` so it is not `None`. As a result, on line 277, the condition is always evaluated as `False`. Is this intended behavior?

https://github.com/fundamentalvision/BEVFormer/blob/66b65f3a1f58caf0507cb2a971b9c0e7f842376c/projects/mmdet3d_plugin/bevformer/modules/transformer.py#L143-L156 https://github.com/fundamentalvision/BEVFormer/blob/66b65f3a1f58caf0507cb2a971b9c0e7f842376c/projects/mmdet3d_plugin/datasets/nuscenes_dataset.py#L78 According to the codes above, the `rotation_angle` used to rotate `prev_bev` is the difference between two patch angles, which is `now_angle - prev_angle`. However, as shown in the...