你好，感谢开源，我仔细阅读代码产生了几个疑惑：
1.在rope3d.py 第 407行代码，tan_roi[h_r, w_r] = (y1_cur - calib.P2[1, 2]) * 1.0 / calib.P2[1,1]，请问这里为什么使用的是calib.P2[1,1]，我理解这里的calib.P2[1,1] 相当于相机内参的fy，与论文中提到的f好像不一致

2. 代码中使用了 min_f = 2250，max_f = 2800，请问这两个数是基于rope3D统计出来的相机焦距范围吗？

3. 在阅读以下代码段时产生了极大困惑，配置文件中interval_max、interval_min的设置依据是什么呢？代码段中的+4.5，-4.5是？
   `if depth_ * _crop_ratio * fp / norm_theat_center3d < self.interval_max[0] / min_f:
   vis_depth[i,0] = (vis_depth_roi* _crop_ratio * fp / norm_theta - (self.interval_min[0]-4.5) / max_f) / ((self.interval_max[0]+4.5) / min_f - (self.interval_min[0]-4.5) / max_f)
   att_depth[i,0] = depth[i] - vis_depth_roi
   depth_bin_ind[i,0] = 1
   depth_mask[i,0] = vis_depth_roi_ind
   ......

            if depth_ * _crop_ratio * fp / norm_theat_center3d >  self.interval_min[4] / max_f and depth_  * _crop_ratio * fp / norm_theat_center3d <  self.interval_max[4]/min_f:
                vis_depth[i,4] = (vis_depth_roi * _crop_ratio * fp / norm_theta - ((self.interval_min[4]-4.5) / max_f)) / ((self.interval_max[4]+4.5) / min_f - (self.interval_min[4]-4.5) / max_f)
                att_depth[i,4] = depth[i] -  vis_depth_roi
                depth_bin_ind[i,4] = 1
                depth_mask[i,4] = vis_depth_roi_ind`
   

十分期待您的答复~

你好，我用像素取点工具取点Cube depth中每个目标的每个面的深度，发现深度值都是整数值，这样的话，cube depth 不会存在很大的误差吗？

同时注意到，该图在使用过程中有一个除以256的动作。
`

rope3d.py第203行

vis_depth_generate_new = cv2.imread('{}/{}.png'.format(self.box3d_dense_depth_dir, index), -1) / 256.`

想请教一下Cube depth在制作过程中是根据公式（6）保存的绝对深度值吗？还是说是计算出z后，会根据公式（4）转换成了Normalized_depth，并且再乘以256形成的值？（目前想要通过自己的数据集制作出cube depth）

Hi, thank you for this great work. I have a question: If I only have an image, can I use the pretrained model to make a prediction without retraining the model with my dataset?

Thanks a lot for your great work! Will you also release the code for Waymo and KITTI? With that I mean configs, command line args, dataset class, .... So everything that is needed to reproduce the paper results.

Hi, there!
Thank you for open-source the code. I am preparing the installation environment.
I notice that：pip install torch==1.5.0 torchvision==0.6.0 numpy==1.23.5 numba==0.58.1 scikit-image==0.21.0 opencv-python==3.4.10.37 tqdm==4.65.0 matplotlib==3.7.1 protobuf==4.22.1 pyyaml==6.0
Since the GPU I am using is 3090, the CUDA for adaptation needs to be above 11.0.
I checked https://pytorch.org/get-started/previous-versions/ . The lowest CUDA11.0 fit for the torch was found to be 1.7.0. What versions of the torch are supported?

Thank you for your reply.

Hello, could you please provide MonoUNI's evaluation indicators for the two categories of Cyclist and Pedestrians on the DAIR-V2x-i data set? I seem to have only seen the category of motor vehicles in the paper：

thanks for your wonderful job and i want to know will you open your model's code ?

Hello,

Is it possible to use other links for the weights and dataset besides Baidu Cloud as it is hard for us outside China to download the files? Thank you very much.

how to train other datasets

The model looks really good and I'd like to experiment with it, but unfortunately I live abroad and don't have a dataset.
So I'd like to try the model with other datasets, and I was wondering if you could tell me when this model will support kitti dataset as well?

您好，十分感谢您的分享，作为一个初学者，我想请问您如下情况是该如何解决：

你好，我训练模型过程中发现 depth_loss 逐渐变成了负数，我发现是loss函数laplacian_aleatoric_uncertainty_loss 中log_variance中出现负值导致的。而log_variance是由一个nn.Conv2d层输出，然后经过torch.logsumexp计算得到， 其中 nn.Conv2d的输出值没有做区间约束，torch.logsumexp计算后就可能存在负值。请问这里地方是否存在bug

I used 4 x TITAN X to reproduce your project.
When I try evaluation, everything goes fine. I got a result like:

However, I tried to train the project several times, and encountered problems differently:

When I use 4 GPU to train the project, the training process stopped with no reason and no error report, it just like:

As you can see, the training ended at 2024-03-29 14:09:12,352 at the fourth epoch, with no feedback on any problems.
The later command to clean the memory with kill -9 was executed by me alone about thirty minutes later, hence this display.
I tried it a few times, and the training always stopped at the first few epochs, taking up GPU memory and only cleaning it up separately.
Before training, I reduced the batch_size to 6. By watching nvidia-smi, I was sure that it was not due to insufficient GPU memory or GPU offline problems.

When I use 2 GPU to train the project, other problems arise, such as:

How to solve these problems, thank you for your advice.

Hello, author!
I enhanced the image of the validation set with a simple rainy & foggy -day simulation program, and it looked like this:

That is, blur the image, reduce the brightness, and add a mask of some raindrops.
I used the official ckpt for verification, and the comparison is as follows:

car
[email protected] [96.97587769060773, 99.43916721862426, 97.04305694184455]
[email protected] [82.91338665086208, 82.22236411339173, 82.18257252897446]
[email protected] [75.5153895932399, 75.27078423603838, 75.1355335795734]
aos [99.27853419324472, 99.37303304437647, 99.33276931153972]
[email protected] [99.45729055841753, 99.5492530942907, 99.51969387573062]
[email protected] [91.80759521645166, 92.49802504973516, 92.47992894413947]
[email protected] [92.65049690746477, 93.06885181584063, 93.0254487356281]
[email protected] [95.01409992120026, 95.54873005854269, 95.53547554066499]
[email protected] [96.40101528049155, 96.5902987969818, 96.57004128525296]
2024-05-21 18:22:22,703
INFO {'[email protected]': [96.97587769060773, 99.43916721862426, 97.04305694184455],
'[email protected]': [82.91338665086208, 82.22236411339173, 82.18257252897446],
'[email protected]': [75.5153895932399, 75.27078423603838, 75.1355335795734],
'aos': [99.27853419324472, 99.37303304437647, 99.33276931153972],
'[email protected]': [99.45729055841753, 99.5492530942907, 99.51969387573062],
'[email protected]': [91.80759521645166, 92.49802504973516, 92.47992894413947],
'[email protected]': [92.65049690746477, 93.06885181584063, 93.0254487356281],
'[email protected]': [95.01409992120026, 95.54873005854269, 95.53547554066499],
'[email protected]': [96.40101528049155, 96.5902987969818, 96.57004128525296]}

car rain and fog
[email protected] [86.62723709493258, 84.31233745769502, 81.93751922630769]
[email protected] [11.97951735346964, 12.91851984096729, 12.953796092913189]
[email protected] [5.602129356343057, 5.8551420736835, 5.849890171969584]
aos [84.02914109959092, 81.92114869458274, 81.86665065264374]
[email protected] [86.99649094524744, 84.62605853292004, 84.59835480523805]
[email protected] [39.08887650358005, 40.64569719077104, 40.62059129107645]
[email protected] [29.25147127531466, 30.762483848887477, 30.714181702464305]
[email protected] [65.57447835387885, 63.23464635696953, 63.1454831542585]
[email protected] [59.41559161753111, 59.2939806263183, 59.16924448602208]
2024-05-21 19:07:44,657
INFO {'[email protected]': [86.62723709493258, 84.31233745769502, 81.93751922630769],
'[email protected]': [11.97951735346964, 12.91851984096729, 12.953796092913189],
'[email protected]': [5.602129356343057, 5.8551420736835, 5.849890171969584],
'aos': [84.02914109959092, 81.92114869458274, 81.86665065264374],
'[email protected]': [86.99649094524744, 84.62605853292004, 84.59835480523805],
'[email protected]': [39.08887650358005, 40.64569719077104, 40.62059129107645],
'[email protected]': [29.25147127531466, 30.762483848887477, 30.714181702464305],
'[email protected]': [65.57447835387885, 63.23464635696953, 63.1454831542585],
'[email protected]': [59.41559161753111, 59.2939806263183, 59.16924448602208]}

In the original val dataset, all APs of car are relatively high at different IOU thresholds. However, after adding the rain-fog mask, all AP decreased significantly, especially the task metrics involving 3D detection.

I ran a similar experiment on BEVHeight and found it much more robust. Does your approach sacrifice robustness for greater accuracy? And any other possible suggestions? Thank you for your reply.

你好👋, 对你的方法非常感兴趣，我想将你们的方法尝试别的数据集, 请问"box3d_depth_dense"这个文件夹里面的深度图是怎么生成的？ 非常感谢