PGD and FCOS3D train with pred_bbox2d=True in config

[betterhalfwzm]

FCOS3D train with pred_bbox2d don't work but PGD works. PGD output result of bbox2d don't process with nms????

[Tai-Wang]

Could you please describe your issue more clearly? What do the "don't work" and "don't process with nms" mean?

[betterhalfwzm]

Could you please describe your issue more clearly? What do the "don't work" and "don't process with nms" mean?

FCOS3D支持pred_bbox2d=True吗？ PGD经过网络得到结果，有一个是bbox2d的结果，这个结果经过nms后处理了吗？

[Tai-Wang]

FCOS3D本身不支持，但应该可以通过微调一些地方也加上相应的模块。 bbox2d的结果应该是没有经过nms后处理的，因为本身主任务还是 3D 检测，加这个 branch 除了 regularize 3D 检测框之外只是为了适配 KITTI 上可以单独输出 2D 框的需求来进行评测，不过这个地方确实还可以优化来提升一下 bbox2d 结果。

[betterhalfwzm]

FCOS3D本身不支持，但应该可以通过微调一些地方也加上相应的模块。 bbox2d的结果应该是没有经过nms后处理的，因为本身主任务还是 3D 检测，加这个 branch 除了 regularize 3D 检测框之外只是为了适配 KITTI 上可以单独输出 2D 框的需求来进行评测，不过这个地方确实还可以优化来提升一下 bbox2d 结果。

谢谢回答！单目检测现在都只能原图进行训练，不能改变resize是吧，然后大概多久会支持这个功能啊，如果自己改代码，需要从哪下手和注意哪些问题，谢谢！

[YinengXiong]

FCOS3D本身不支持，但应该可以通过微调一些地方也加上相应的模块。 bbox2d的结果应该是没有经过nms后处理的，因为本身主任务还是 3D 检测，加这个 branch 除了 regularize 3D 检测框之外只是为了适配 KITTI 上可以单独输出 2D 框的需求来进行评测，不过这个地方确实还可以优化来提升一下 bbox2d 结果。

谢谢回答！单目检测现在都只能原图进行训练，不能改变resize是吧，然后大概多久会支持这个功能啊，如果自己改代码，需要从哪下手和注意哪些问题，谢谢！

Looking forward to Resize3D too，cam2img should be changed when resizing image size And also I have another question: when I was training SMOKE model on KITTI dataset, I found RandomShiftScale could boost my performance, but when I change to nuScenes dataset, adding RandomShiftScale to my data_pipeline will decrease my performance a lot, what causes this problem? besides, when using RandomScale, is it necessary to change the GT depth accordingly?

[Tai-Wang]

1. Resize3D and more pipelines will be supported in a month.
2. I did not try RandomShiftScale in SMOKE, but Resize3D may not bring gains on nuScenes from my experience. It might be related to different focuses of KITTI and nuScenes metrics.
3. From my experience, only rescaling 2D attributes is enough most of the time. If you would like to change GT depth to make the rescaling also take effect on 3D attributes, it is necessary to change other 3D targets to make the geometry consistent. From our attempts, it may not bring much performance difference on KITTI.

[YinengXiong]

1. Resize3D and more pipelines will be supported in a month.
2. I did not try RandomShiftScale in SMOKE, but Resize3D may not bring gains on nuScenes from my experience. It might be related to different focuses of KITTI and nuScenes metrics.
3. From my experience, only rescaling 2D attributes is enough most of the time. If you would like to change GT depth to make the rescaling also take effect on 3D attributes, it is necessary to change other 3D targets to make the geometry consistent. From our attempts, it may not bring much performance difference on KITTI.

For 2. my nuScenes dataset is converted to KITTI format, and RandomShift can improve the performance on KITTI while the KITTI format nuScenes dataset is the opposite

[erwangccc]

@Tai-Wang Hi, Thanks for your great work. When will the 3D resize feature be released? Looking forward to this feature. Thanks. @betterhalfwzm Hi, have you tried to resize the data of nuscenes to train FCOS3D normally? Thanks.

[Hukongtao]

1. Resize3D and more pipelines will be supported in a month.
2. I did not try RandomShiftScale in SMOKE, but Resize3D may not bring gains on nuScenes from my experience. It might be related to different focuses of KITTI and nuScenes metrics.
3. From my experience, only rescaling 2D attributes is enough most of the time. If you would like to change GT depth to make the rescaling also take effect on 3D attributes, it is necessary to change other 3D targets to make the geometry consistent. From our attempts, it may not bring much performance difference on KITTI.

Hello, Is Resize3D now?