SuperPoint discussions - Githubissues

27/09/2024

[x] Extract array of points from pretrained version. What's the difference in the generation of descriptors and the features extracted by the network?
[x] Check if it's possible to directly use the pretrained method in order to avoid needing to fine-tune/train from scratch on different datasets
[x] Check the tuning of SCARF parameters
[ ] Check if the trainable version has some released weights
[ ] Confront the two methods based on accuracy and robustness
[ ] If possible confront the model running on TensorFlow with the one running on Pytorch
[ ] Try to run on edge detectors when it will be needed to compare SCARF with RGB images
[x] Undestand the need for fine-tuning/training and which datasets can be used for training

[x] Check how descriptors are created, matched and how to access them
[x] Try out OnePose with frame-based sequences
[ ] Try out OnePose with SCARF sequences
[x] Check the sequence with a real event-camera (events and SCARF) and YCB object
[x] Populate the dataset keeping in mind that the system will use a real camera and will scan the object going around it
[ ] Find good tuning values for SCARF based on the new sequence
[x] Try SP on new sequence