wangzhaoming

Results 15 comments of wangzhaoming

> 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?

> 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!

I have met this problem. Anyone fix it ? @javierjsa @robotzheng @javierjsa @dajiangxiaoyan @realwecan

I have the same problem,ImportError: No module named cpu_nms/gpu_nums #66;Do you know how to fix this? Thanks a lot! @lnuchiyo

> 在 RandomCrop 之前加同样的 resize 即可。 @Tai-Wang 训练的pgd,跑mono_det_demo.py时,3D框位置不对 ![image](https://user-images.githubusercontent.com/20449990/186342557-da1f2ae7-ada3-405b-8b47-887b5d7c31df.png) ![2cac95007c22ed765c25340f1549262](https://user-images.githubusercontent.com/20449990/186316451-4d780654-d80f-4097-a017-ce6fc3c69e0b.jpg) ![f6c35b8d36e507c87b9450e7989e307](https://user-images.githubusercontent.com/20449990/186316486-bc8a539a-7e8d-42ba-938f-5206496c8c1c.jpg) 训练ap: ----------- AP11 Results ------------ Pedestrian [email protected], 0.50, 0.50: bbox AP11:16.3295, 8.1563, 8.1664 bev AP11:9.0909, 9.0909, 9.0909 3d AP11:9.0909,...

> 请问您用的哪个config?看起来不像是repo里给定的。FCOS3D或者PGD在KITTI上的Resize3D如何使用还有待测试,如有需要可以先用DfM和Waymo上PGD的config尝试。 @Tai-Wang pgd kitti那个config,换了个backbone,其他的应该一样 我尝试了训练原图,测试640*384,3d位置不对; ![image](https://user-images.githubusercontent.com/20449990/186380802-99770aa1-5dac-45b7-863b-05b515223ed2.png) 训练图片640*384,测试640*384,3d位置不对; ![image](https://user-images.githubusercontent.com/20449990/186379571-c34fda64-1ba8-4929-b305-8435dc214927.png) 训练图片640*384,测试原图,3d位置不对; ![image](https://user-images.githubusercontent.com/20449990/186380010-29d8ed7b-87da-4c9c-bbfd-bfb80aee2551.png) 只有原图训练原图测试3d位置才对的上 ![image](https://user-images.githubusercontent.com/20449990/186380548-b326710e-556a-42bf-95dc-a239b7516d66.png)

当scale_factor=0.515时,是这个效果,训练图片640*384,测试640*384,kitti的原图是1242, 375,必须保持同比例缩放吗? ![9f8e11949dd5887decf327153706185](https://user-images.githubusercontent.com/20449990/186393357-53795e63-ea6d-4bc7-91d5-7f944dcbb38c.jpg) ![image](https://user-images.githubusercontent.com/20449990/186393429-9f396fd4-7d55-4c1d-b578-ee69d6ee127c.png)

pgd 训练nus,@Tai-Wang ![image](https://user-images.githubusercontent.com/20449990/186871705-ff48dc0d-1e78-43b2-ba0d-f2fd94515369.png) train: ![image](https://user-images.githubusercontent.com/20449990/186872018-84cc5e2f-cbe0-4562-9809-5ca0b11218fb.png) test: ![image](https://user-images.githubusercontent.com/20449990/186871957-d78cf02b-37e5-4c3a-ac36-d6c01f1ea6c4.png)