search

csiro-robotics / InCloud

[IROS2022] Official repository of InCloud: Incremental Learning for Point Cloud Place Recognition, Published in IROS2022 https://arxiv.org/abs/2203.00807
https://ieeexplore.ieee.org/document/9981252
Other
36 stars 9 forks source link

about Fig. 4 in your paper #2

Closed BIT-MJY closed 1 year ago

BIT-MJY commented 1 year ago

Hi authors,

Thanks for your fantastic work!

I have just read your paper and found that in Fig.4 all three PR methods perform better on Oxford after being trained on Riverside, which seems like the opposite trend to catastrophic forgetting. Could you please explain the results?

peymmo commented 1 year ago

@JBKnights

JBKnights commented 1 year ago

Hello, thanks for the question!

Our experience producing this paper was that the Riverside environment has stronger transferability to Oxford than the DCC environment. The increase in performance as a result of training Riverside is not indicative of "remembering" performance on Oxford in any way, but instead reflective of the model overfitting to the Riverside environment - an environment with greater transferability to Oxford than DCC, the last previously overfit-to environment.

Applying incremental learning to this scenario reduces this volatility by preventing the forgetting and overfitting from taking place, maintaining reliably high performance on the previously visited environments. I hope this answers your question!

BIT-MJY commented 1 year ago

but instead reflective of the model overfitting to the Riverside environment - an environment with greater transferability to Oxford than DCC

Really insightful answer! I will close this issue.