RMSE ATE will suddenly become much larger

[leqwq103]

Hello! Recently, while practicing on MonoGS code, I discovered an issue. During the operation, I, MonoGS: Evaluating ATE at frame: 1025 Eval: RMSE ATE [m] 0.010714135020835631 MonoGS: Evaluating ATE at frame: 1121 Eval: RMSE ATE [m] 0.014316433431372504 MonoGS: Evaluating ATE at frame: 1200 Eval: RMSE ATE [m] 0.02489313051194339 MonoGS: Evaluating ATE at frame: 1286 Eval: RMSE ATE [m] 0.031591810183650405 RMSE ATE will suddenly become much larger. I checked the saved images and found that there is a curve in the map where the RMSE ATE value will increase. I would like to seek your help and see if there is a good solution?

[leqwq103]

In the data set of rgbd_dataset_freiburg3_long_office_household, is the average RMSE ATE of MonoGS 0.0439m? Similar to the final output: Eval: RMSE ATE [m] 0.0439...

[muskie82]

Hi, It's understandable that some parts of the sequence, such as around the teddy bear in fr3/office due to strong rotations and close-ups, might show higher error. As long as the final ATE RMSE number is reasonable, I think it is OK.

Best,

[z157181773]

I also have a question about RMSE ATE because I am using Python slam.py -- config configs/mono/tum/fr3_office. yaml -- eval. Regarding mono/fr3 in the paper, the table in the paper states that RMSE ATE is 4.39cm. I would like to know if this is the average?After I improve the code, can I compare it with this value?

[muskie82]

It is the average number of three runs as mentioned in the paper. Yeah, just compare to it.