How to fine-tune DPVO on your own dataset and achieve better results？

[JingruiYu]

Hello, Great work! so I tried it on my own dataset, but the results are quite strange.
I prepared the data in the format of TartanAir. Under each scene folder, there are depth_left folders, image_left folders, and pose_left.txt file.
The format of the pose_left.txt file is tz, tx, ty, qz, qx, qy, qw. I noticed that the code adjusts the columns when loading the pose txt file.
Additionally, the resolution of my images is different from that of TartanAir, so I modified the code to read the intrinsic parameters from a different location.
After preparing the data, I used the code like here to generate the pickle file.
BTW, I removed the operation poses[:,:3] /= DEPTH_SCALE, thinking it wouldn't have a significant impact.
Then, I used a train command almost identical to the one in the readme, except that I reduced the --lr to 0.00004 for fine-tuning. I mixed my own data with the TartanAir data for this process.

However, the trajectory results plotted during the fine-tuning process are quite poor. For example, they look like this.
step=8000

step=16000

step=30000

But before fine-tuning, the results obtained using the downloaded model were like this.

In other words, the results have become worse. Under the same training steps, the results on the TartanAir data were perfectly normal.

I'm not sure where I might have gone wrong or what the reason could be. Thank you!

[akramsalim]

How did you prepared your dataset in Tartan format? Was the dataset RGB and Lidar then your converter them into RagB-D ? Or how did you do it?