Coming soon.

Sorry! This work is private yet. The public version will release soon. Stay tuned!

Nvidia OptiX is a promising GPU-based ray-tracing SDK.

I am currently implementing an OptiX-based program that automatically generates rendered images especially for training deep convolutional denoising models.

The program will involve random camera transformations, random material changes, auxiliary buffer processing, etc.

In this paper, the authors propose a two-headed neural network and maximize the discrepancy between the two classifiers to detect out-of-distribution samples while correctly classifying in-distribution inputs. Also, they suggest a new problem setting where they utilize unlabeled data for unsupervised training, whereas previous works only exploit labeled in-distribution samples.

In this paper, the authors propose a task-agnostic and straightforward active learning technique. They build a “loss prediction module” that predicts target losses of unlabeled inputs. Using this predicted loss as an uncertainty measure, they determined which of the unlabeled data would be most valuable to annotate.