cvg
the matching between day and night
Hi,Thx for your great work..
I found the matching between day and night is bad,can you give me some suggestion? Than you very much. I use the pre_trained model for test and also train by myself.
Hi, the pre-trained model was not trained on day-night images and is thus not robust to it. If you want to make it more robust to these changes, I would suggest training on another dataset or adding strong photometric augmentation to mimic night-time images. For the day-night dataset, you could try the Aachen dataset like in the R2D2 paper, or the VIDIT dataset for example.
Hi, the pre-trained model was not trained on day-night images and is thus not robust to it. If you want to make it more robust to these changes, I would suggest training on another dataset or adding strong photometric augmentation to mimic night-time images. For the day-night dataset, you could try the Aachen dataset like in the R2D2 paper, or the VIDIT dataset for example.
Hi,thanks for your reply. I have trained new model in the dark dataset. Between two frames(day vs day or night vs night),the matching looks reasonable,but the day-night matching looks unreasonable.According to my understanding, using unsupervised model does not train descriptor during the day and night,the result maybe not very good because of this.
I am not sure to understand correctly. Did you have day-night pairs in your training set or only day-day and night-night pairs? Because you would need to train with day-night as well. If you did, then I am afraid that there is not so much that you can do. Matching across large illumination variation is still an ongoing direction of research nowadays.
Sorry, I didn't make it clear. I trained the dataset include the day images and night images,but not including day-night pairs because I didn't know the pre-matching relationship between them. When use the own trained model for test,Between two frames(night vs night),the matching looks reasonable,but the day-night matching looks unreasonable.
I see, then there is hope to improve your day-night matching performance. You need to have day-night pairs in the training set as well, for example using VIDIT, where it is easy to get the ground truth matching.
Hi, the pre-trained model was not trained on day-night images and is thus not robust to it. If you want to make it more robust to these changes, I would suggest training on another dataset or adding strong photometric augmentation to mimic night-time images. For the day-night dataset, you could try the Aachen dataset like in the R2D2 paper, or the VIDIT dataset for example.
I am closing this for now, feel free to reopen if you need additional assistance.