Xinpeng Liu

Yes, and we optimized the thresholds for different object categories considering the frequency in the whole dataset.

It is because the inter_det_mask and nis_mask are using different detections. If you are using customized object detection result, you might need to also generate a customized NIS file (to...

You could try to save these two predictions for visualization. https://github.com/DirtyHarryLYL/DJ-RN/blob/83c3d0f16a4a3b55545e96a8350397f6adb1ef91/lib/networks/DJR.py#L550-L551

We re-struct the official `.mat` format annotation into the provided file. If needed, we would update the re-struct script.

Please take reference from the following code. For convenience, we incorporated the calculation of f_{H}^{3D} in the full model. https://github.com/DirtyHarryLYL/DJ-RN/blob/83c3d0f16a4a3b55545e96a8350397f6adb1ef91/lib/networks/DJR.py#L383