serenos / lwsis Goto Github PK

When I reproduced CondInst, the results were different from those provided in the paper.
I train the ResNet50-Condinst for 90K iterations with batch size 16 on 4 3090 GPUs.
The results were as follows:
AP 32.015 AP50 57.328 AP75 30.931
CAR 43.365 Ped 33.444 Cyc 19.237
According to my observation, there is no instance id in waymo dataset with bicycle and ped object.

When I open adet/modeling/condinst/dynamic_mask_head.py and adet/modeling/condinst/register_lidar_annotations.py on Windows, I find that these two files are corrupted:
%TSD-Header-###%�ewzc{Iv��t�+��N� � @ � m�%x�t 8�P ����������~[H�4�6����73�⛌@��߮���V���R�������F;��as�����?�Q>J'��0Eߔ7q�a?��R��������~[H�4������72�G��@K�߮(��V���R�������~�4�oc������ڢ��br�HH��� F�����73����>���ga4y�Fo0Ȝ���0��͗Z���=�8�������Y��K���6A����p�����ܞ,$'�A�Xi��
Ӏx�T���х1g�X����O&q���$8���ˁ�q��>U((��3����4�(�t{�MA5?������V?O�k�g��� ���xeAQ�i�~y���bl��Y�lm���E!���0Ft�֯_�O������������������T�/��;���t��#�w�V���A��e�J���2��^#�����lF2.����F��у�l����Y�Djm��黄��K4���m>!vy��r��}̸���<N�6��E�h^Q����T��� �E"�JX�̹F��y�=�"��:�������f�������J�f��2��t*:��bϊQ9+��[�I�7g�����(���h?�J\�lI��D�������%vݭ1��C�����w�yο�]{��;ScB8Y6��{adi�r��4��M�Q>�������Oў��n7����$�-�l�q��ݍ�w��9L��X��L���\��ӇD͚���Q���e�ɐ����\���P'��O\�UG,=���3���!�;@j�W�Gϒ���(?��ǀ�cM�Y.<�����S�8���3'���e�G��4�2���������*\� �nFȒ��T��_��.݁AZc�J�|�@�q W�d�OL� Z�k :Ӱ�����>՚�����������pjv�{�g0 �s������wSoqSgyY�#�c�x��]!�b��ƭ�S<r[��C��B����6ߜ�����{�Z��z�oS����@\���o��;9������)���l���w��ޣk��_��tx�5b�/�m�B�Q�� ~]�/����.Ѯ�9������"o��R������b��u�Ң���|���0�P�i}@��ND�HŨ+}�0 ��6m�V�������9EJ���^���0(�R���0n�%��D�h:v.FD�n�W���s=�'O���*nB�������Yz��?�S�����]�����ҐG�l�4��%C�����h�L�(O�(�'W.�OB9Z#ogI�z�=�I���������h 0k�[ ��������?��Sc�wi����zT9���Bx�r�B�PY'@�C�;E�>��9��/c����B?�Cm&�������7��K�u�� ����� ʅ���ֽ��9�*����m���,�6!t\ !��h�|F]�eE������Ԏ����J��
��]�$���/t��L�W+|PL4h%9�F�i���`

Thank you for you great work!
May I ask when the code will be released?

How do you handle waymo instance annotations? I use waymo dataset to train the condinst and boxinst on mmdet2.x. But boxinst's mAP is higher then condisnt mAP.I suspect I mishandled the instance annotations.