tyxsspa

demo内置的翻译模块大约消耗4G显存，如果需要使用使用CPU推理，因官方尚未支持，需要手动修改modelscope中翻译模块的推理代码，通常位于如下类似位置(仅举例)： `/home/username/anaconda3/envs/anytext/lib/python3.10/site-packages/modelscope/pipelines/nlp/translation_pipeline.py` 然后在TranslationPipeline的__init__()函数中添加如下代码： ``` ... tf_config.gpu_options.allow_growth = True tf_config.device_count['GPU'] = 0 #

why all 12 views use same aug?

1

hi, I notice that you use line 384 instead of line 388 https://github.com/megvii-research/PETR/blob/7b046e28a03dcb296987eb3e76cbf761f52a37f0/projects/mmdet3d_plugin/datasets/pipelines/transform_3d.py#L384-L390 which will use same aug in all 12 views, I wonder why? Have you test that different...

Maybe pc_range not match in 3DPE and bbox_pred?

1

the pseudo points are in range [-61.2, -61.2, -10.0, 61.2, 61.2, 10.0] when doing 3D-position-embeding, then normalized to (0,1) and inverse_sigmoid. query_embs are also in (0,1) and inverse_sigmoid, then query...

rotate_bev_along_z has bug？

1

May be the extrinsics after bev rotate here https://github.com/megvii-research/PETR/blob/7b046e28a03dcb296987eb3e76cbf761f52a37f0/projects/mmdet3d_plugin/datasets/pipelines/transform_3d.py#L669 should be exactly same as in https://github.com/megvii-research/PETR/blob/7b046e28a03dcb296987eb3e76cbf761f52a37f0/projects/mmdet3d_plugin/datasets/pipelines/transform_3d.py#L688 because you use camera extrinsics as **transposition** transform matrix from L to C, so...