Yifan Zhang
Yifan Zhang
The issue could be due to the PyTorch version or jitter caused by random sampling of input point clouds.
My code is based on specific version of pcdet, and we do not test the code on other versions.
What's the error log information when you run the attack.py directly?
The BAFT implementation for adversarial point perturbation is here: https://github.com/Eaphan/Robust3DOD/blob/main/tools/train_utils/train_utils_BAFT_perturbation_demo.py Please change the hyper-parameters according to the type of detectors you use, such as "key" and "model_name". Then use the...
We use the mAP, which is the average AP@R40 0.7 0.7 0.7 (3d) on three classes for kitti dataset. mAP ratio = mAP under attack/ mAP without attack.
Sorry for the confusion. AP@R40 0.7 0.7 0.7 (3d) for car AP@R40 0.5 0.5 0.5 (3d) for pedestrian and cyclist. Then mAP is the average of them (9 values). If...
1. 涉及训练速度的要素太多，根据你提供的信息我无法帮你定位原因. 2. 前几个epoch我们先用特征来更新vMF分布，几个epoch之后才加入regularization (KL) loss. 3. 我们提供的是预训练权重，但是这个权重无法直接用于语义分割任务，也就是不具有zero-shot的能力. 你需要的功能可以看看Clip2scene 这个文章.
Please save the adversarial point clouds. Then feed them to the target model and do not use voxel-sampling for the pre-processing.
I created the labels using the code in the demo_code directory. I'm unable to explain the difference.