confusion about model
ds22058 opened this issue · 4 comments
Hi, what I understand is that ActiveNeRF selects the best subset of samples for the model from a large number of samples, but in this way, it still need a larger number of samples first. In practical applications, it may only have less than ten images for a scene, and in this case, ActiveNeRF cannot be used. I wonder if I misunderstand the model. Secondly, compared to NeRF, ActiveNeRF only uses a small number of samples to train and ultimately achieves similar results. Can it be considered that its feature extraction ability and model fitting ability are stronger, similar to those models with feew-shot input. Looking forward to your reply!
Hi! Thanks for your attention to our work.
For the first concern, our example for lego or hotdog keeps 20 initial training samples while other 80 images are viewed as a holdout set from where we actively capture new data in the active iteration. It is also possible to set the number of initial training samples to a smaller value, such as 10 or 5.
For the second question, we find that modeling NeRF with uncertainty can in fact bring improvements under low-data regime. Nevertheless, there are also other approaches that target on scene/object reconstruction under extremely less data (e.g., 1), and in that case uncertainty alone is not enough. Comparably, ActiveNeRF is good at improving reconstruction ability with minimum additional resource, if new data can be captured in the training process.
@Panxuran
When initially select the training camera poses, I noticed that it chooses interval selection, and not according to the spatial distribution of the selection, so will not be too strict ah?
@YZsZY
The data choice in our example is only a toy illustration. You can choose your own initial training set that most suitable for your method or task.
Hi, we want to reproduce the results in your Table 2. From the caption, Setting I is initally trained with 4 observations. I wonder how you choose these 4 observations. Thank you !