Question about the smaller model #112
Description
Thank you very much for sharing this outstanding work. I would be most grateful if you could clarify one point regarding the “(ii) Smaller model” in Section 4.4 Ablation. Specifically:
- Did you halve the number of channels in the last three convolutional blocks for all of the Keypoint Heatmap, Descriptors, and Reliability Map to 32? Or were there any additional modifications or tricks involved?
I have tried to reproduce this smaller model by halving those channels as illustrated in the figure below, but obtained the AUC@5 of Xfeat only around 33 (rather than the 37 reported). Could you kindly provide more details on how the (ii) Smaller model was implemented?
