Exploring CLIP Feature Vectors for Improved Out-of-Distribution Detection
Jessica Li · 2023
Out-of-distribution detection in the realm of computer vision is dominated by deep generative models usually trained on raw pixel images. The results of out-of-distribution detection with generative models vary. Among the generative models used today for out-of-distribution detection, diffusion models seem to perform well however, diffusion models tend to be computationally expensive and slow. In an effort to improve out-of-distribution detection using generative models, we introduce a diffusion based autoencoder model. Another area of interest is the use of OpenAI's Contrastive Language-Image Pre-training (CLIP) image encoder to create CLIP feature vectors of our datasets for use with our model. We compare the performance results of both the use of CLIP feature vectors and the use of raw pixel images with our model. We also test the performance of our model against benchmark results of other models such as a diffusion model.