Frequently Asked & Good Questions

[haibao-yu]

We will continually gather both frequently asked- and good questions in this thread.

Frequently Asked Questions

1. How to use dataset outside mainland China? Dataset request Outside Mainland China, Request access for DAIR-V2X
   Answer: At present, the dataset is officially available for download and use exclusively within Mainland China. To help researchers become acquainted with our data, we have provided a Google Drive link for direct download of an example dataset. If you wish to utilize our complete dataset, we encourage collaboration with researchers based in Mainland China.

   • DAIR-V2X-C-Example: google_drive_link
   • V2X-Seq-SPD-Example: google_drive_link
   • V2X-Seq-TFD-Example: google_drive_link

2. Why there are two kind of labels, and how to choose label for multimodal fusion detection? Two kind of labels and multimodal fusion model
   Answer: There is no strict synchronization between the camera and LiDAR. Consequently, we supply two types of labels, each corresponding to either the camera or LiDAR data. For effective multimodal fusion detection, we recommend utilizing LiDAR as the primary sensor, with its corresponding label serving as the ground truth.

3. The evaluation results are not same as the official results while following the tutorial.
   Answer: There are several common reasons for issues encountered. Firstly, it is essential to ensure that the tutorial has been followed accurately; kindly review your process to identify any discrepancies. Secondly, be mindful of cache files as they can impact the running results. Specifically, certain cache files, such as './cache/tmps,' may not be updated after generation, as this is done to optimize the inference speed.

4. The rotation rotation matrix from virtual lidar to camera is not standard rotation matrix. Question about rotation matrix from virtual lidar to camera
   Answer: We conducted a secondary calibration to generate a new transformation matrix, aiming to rectify errors present in the original calibration. The oversight of omitting constraints led to a non-standard rotation matrix. This issue can be resolved through the generation of a standard rotation matrix approximation.