Closed ghoshaw closed 4 years ago
Hi @ghoshaw, thanks for your interests in our work. The metrics used in our paper are the same as the ones in VLocNet++. VLocNet++ makes a good combination of relocalization, odometry and semantic segmentation based on ResNet-50. Similarly, [1] also jointly regresses the global pose as well as the relative pose using ResNet-34, which just achieves 10 times worse results than VLocNet. From these, multi-task learning and deeper architecture indeed boost the performance. You are encouraged to make improvements on localization-related tasks form these two aspects.
[1] Xue, Fei, et al. "Local supports global: Deep camera relocalization with sequence enhancement." ICCV 2019.
ok, thanks for your reply!
HI, thanks for your sharing, but I have some questions about the result of the pose regression method. For A paper "VLocNet++: Deep Multitask Learning for Semantic Visual Localization and Odometry" published in 2018.4.24, their performance is pretty good, as least in the paper and they seems not have much fancy things. And by now, no offence, used attention etc, the result on 7-senses seems 10 times less. I just get confused, Is that the metrics you used different or something else?