Artificial Intelligence in Computer Vision: A Study Beyond ImageNet Accuracy
There has been a significant increase in the complexity of computer vision models in recent years. From ConvNets to Vision Transformers, there are now many models available for use. With the shift from supervised learning on ImageNet to self-supervised learning and image-text pair training, such as CLIP, new training paradigms have emerged.
A new study by MBZUAI and Meta AI Research is examining model characteristics beyond ImageNet correctness to better understand the behaviors of vision models. The study explores the distinct advantages of CLIP’s visual encoder and the model properties of ConvNeXt, ConvNet, and Vision Transformer trained using supervised and CLIP methods.
Through this research, the team aims to shed light on the intrinsic qualities of vision models and discover how they perform in practical scenarios. Their findings highlight the need for new, independent benchmarks and comprehensive evaluation metrics for precise model selection.
The study recommends the use of supervised ConvNeXt for tasks similar to ImageNet and CLIP models for significant domain transitions. The results illustrate that different models show their strengths differently, and a single statistic cannot adequately measure these differences.
For more information, refer to the study’s Paper, Project, and Github, and follow the researchers on Twitter. If you’re interested in receiving more updates on AI, subscribe to their newsletter and join their Telegram Channel.