Ashish Patel 🇮🇳’s Post

𝗗𝗮𝘆-𝟱𝟬𝟬 𝗖𝗼𝗺𝗽𝘂𝘁𝗲𝗿 𝗩𝗶𝘀𝗶𝗼𝗻 𝗟𝗲𝗮𝗿𝗻𝗶𝗻𝗴 Deep Spectral Methods: A Surprisingly Strong Baseline for Unsupervised Semantic Segmentation and Localization by University of Oxford Follow me for a similar post: Ashish Patel ------------------------------------------------------------------- 𝗜𝗻𝘁𝗲𝗿𝗲𝘀𝘁𝗶𝗻𝗴 𝗙𝗮𝗰𝘁𝘀 : 🔸 This paper is published ARXIV2022. 👉 Introduced a method for unsupervised localization, segmentation and matting based on spectral graph theory and deep features. 👉 Despite the simple formulation, it achieves state-of-the-art unsupervised performance for these tasks. It is interesting to note that this performance is only achieved with features from transformer architectures and not with CNNs, which we attribute to the inherent functionality of self-attention in transformers that aligns well with dense localization tasks. 👉 While spectral graph theory has been relegated to a minor role in the age of deep learning, we find that the inductive biases on which it is built can be very useful in the unsupervised setting. 📰 Paper : https://lnkd.in/d3XS5aYk ------------------------------------------------------------------- 𝗜𝗠𝗣𝗢𝗥𝗧𝗔𝗡𝗖𝗘 👉 Unsupervised localization and segmentation are long-standing computer vision challenges that involve decomposing an image into semantically-meaningful segments without any labeled data.  👉 These tasks are particularly interesting in an unsupervised setting due to the difficulty and cost of obtaining dense image annotations, but existing unsupervised approaches struggle with complex scenes containing multiple objects.  👉 Differently from existing methods, which are purely based on deep learning, we take inspiration from traditional spectral segmentation methods by reframing image decomposition as a graph partitioning problem. 👉 Specifically, we examine the eigenvectors of the Laplacian of a feature affinity matrix from self-supervised networks.  👉 We find that these eigenvectors already decompose an image into meaningful segments, and can be readily used to localize objects in a scene.  👉 Furthermore, by clustering the features associated with these segments across a dataset, we can obtain well-delineated, nameable regions, i.e. semantic segmentations.  👉 Experiments on complex datasets (Pascal VOC, MS-COCO) demonstrate that our simple spectral method outperforms the state-of-the-art in unsupervised localization and segmentation by a significant margin. 👉 Furthermore, our method can be readily used for a variety of complex image editing tasks, such as background removal and compositing. #computervision #artificialintelligence #deeplearning