After retraining, I found that the projection results always shift forward. If the official model is used, it is normal. Has the author encountered it? about bevdet

[abing222]

map and nds are normal, bev projection is normal, and z prediction is poor

[abing222]

There is no problem with the visualization code. The visualization of the official model is completely normal. Such problems will occur with the retrained model

[HuangJunJie2017]

It seems ok when I use the '--show' in tools/test.py. The predicted bounding boxes of BEVDet are tightly surrounding the lidar points of the targets.

[abing222]

It seems ok when I use the '--show' in tools/test.py. The predicted bounding boxes of BEVDet are tightly surrounding the lidar points of the targets.

There is no problem with Bev, but the accuracy of Z cannot be seen

[abing222]

It seems ok when I use the '--show' in tools/test.py. The predicted bounding boxes of BEVDet are tightly surrounding the lidar points of the targets.

The officially provided model visualization is normal, but the Z error will be large after retraining, but this will not affect the map and NDS

[HuangJunJie2017]

@abing222 The z prediction seems ok in my visualization.

[abing222]

@abing222 The z prediction seems ok in my visualization.

There is no error in the provided model, and Z prediction is normal, but Z will have a large error if the model is retrained

[HuangJunJie2017]

@abing222 I got it, you mean there is something wrong with the training process. I will check this.

[abing222]

@abing222 I got it, you mean there is something wrong with the training process. I will check this.

thank you

I have checked the bbox and camera external parameters after Bev enhancement, and there is no problem, that is, GT is no problem

[HuangJunJie2017]

@abing222 What configuration do you use for retraining, bevdet or the bevdet4d?

[abing222]

@abing222 What configuration do you use for retraining, bevdet or the bevdet4d

bevdet tiny

[HuangJunJie2017]

@abing222 What configuration do you use for retraining, bevdet or the bevdet4d

bevdet tiny

got it

[HuangJunJie2017]

@abing222 may I have your training log? [email protected]

[abing222]

@abing222 may I have your training log?

@abing222 may I have your training log? [email protected]

yes

[MandyMo]

同样碰到了类似的问题，不知道作者有没有碰到，嘻嘻。

[HuangJunJie2017]

@MandyMo It seems like some false positive result, i suggest you to use a higher score threshold~