The ATE results is different from the paper.

[AndreGao08]

Excellent work! I have followed the instruction and run the whole project in TUM-VIE Dataset, but the ATE results is different from the papar. Even Better Than Paper!!!

Could you help me figure out why？

[huangjian39]

Thanks for your interest in our work, and great to hear that you were able to run the system on the TUM-VIE dataset!

Regarding the difference in ATE results: IncEventGS is an iterative optimization-based method, so the final trajectory can vary slightly across runs. The results are also influenced by the initial depth estimation during initialization.

In the paper, the reported ATE values were obtained by running the system three times and taking the average RMSE. Therefore, it is possible that a single run on your side yields slightly better or slightly worse results than the paper’s averaged numbers.

[AndreGao08]

I want to train the whole sequence. When I set the training_start_index as 0, and training_end_index as -1. It came with error. Do you know why?
Exception has occurred: IndexError
index 50 is out of bounds for axis 0 with size 50
File "/home/sim/Documents/IncEventGS/main.py", line 991, in initialization
if(incre_sampling_segs_end[start_tmp_]<=max_n_winsize):
File "/home/sim/Documents/IncEventGS/main.py", line 2575, in run
self.initialization(init_batch, niters=self.config['num_opti_steps_for_depthEst_init'], traj_mode='linear', depth_map=None)
File "/home/sim/Documents/IncEventGS/main.py", line 2980, in
slam.run()
IndexError: index 50 is out of bounds for axis 0 with size 50

Could you please clarify what specific training_start_index and training_end_index values were used to generate the quantitative results (especially the time/performance metrics) in the paper? Did you utilize the entire sequence, or was a specific subset used to avoid this initialization skipping?

Any guidance on the correct configuration for reproduction would be greatly appreciated.

[huangjian39]

It looks like this error:

index 50 is out of bounds for axis 0 with size 50
File "/home/sim/Documents/IncEventGS/main.py", line 991, in initialization
    if (incre_sampling_segs_end[start_tmp_] <= max_n_winsize):

is caused by having too few events at the beginning of the sequence. We don’t recommend starting training from index 0, since the camera is still initializing during the first moments and the event count is very low and unstable. Also, the data used in our paper were all run with the default JSON files provided in the repo. We don’t use the entire sequence; evaluation is done on a specific subset.

[AndreGao08]

@huangjian39 Thanks for the clarification! I understand now why starting from index 0 causes the error due to the unstable initialization of the camera.

Following your advice, I checked the default configuration file in the repo and found these settings in configs/TUM_VIE/mocap-desk2.yaml:
training_start_index: 50 training_end_index: 90 exp_name: demo
I would like to double-check regarding the paper's results: Did the quantitative evaluation in the paper use this specific short range (50 to 90)?

[huangjian39]

yes