Hierarchical Compact Clustering Attention (COCA) for Unsupervised Object-Centric Learning

TL;DR

The paper addresses unsupervised object-centric learning from single images by introducing COCA-Net, a hierarchical clustering-based encoder built around the Compact Clustering Attention (COCA) layer. COCA integrates a HAC-inspired clustering process with a compactness-based objective to identify object centroids and generate object masks without predefined slot counts, yielding strong decoder and encoder segmentation performance. The approach supports dynamic slot allocation, robust background handling, and high-quality encoder features, demonstrated across six synthetic datasets with favorable results and ablations showing the value of the compactness criterion. The work advances unsupervised OCL by providing a scalable, interpretable hierarchical architecture and paving the way for encoder-based object-centric feature extraction in downstream tasks.

Abstract

We propose the Compact Clustering Attention (COCA) layer, an effective building block that introduces a hierarchical strategy for object-centric representation learning, while solving the unsupervised object discovery task on single images. COCA is an attention-based clustering module capable of extracting object-centric representations from multi-object scenes, when cascaded into a bottom-up hierarchical network architecture, referred to as COCA-Net. At its core, COCA utilizes a novel clustering algorithm that leverages the physical concept of compactness, to highlight distinct object centroids in a scene, providing a spatial inductive bias. Thanks to this strategy, COCA-Net generates high-quality segmentation masks on both the decoder side and, notably, the encoder side of its pipeline. Additionally, COCA-Net is not bound by a predetermined number of object masks that it generates and handles the segmentation of background elements better than its competitors. We demonstrate COCA-Net's segmentation performance on six widely adopted datasets, achieving superior or competitive results against the state-of-the-art models across nine different evaluation metrics.

Figures (11)

• Figure 1: Compactness scores obtained for each pixel in the scene, across four different datasets. The transition from bright yellow to deep purple signifies decreasing compactness. To obtain these scores, a trained COCA-Net encoder is used to generate object masks. Each object mask is then broadcasted to pixels based on the pixel-object assignments. This operation associates every pixel with a copy of its object's mask. Finally, compactness scores for each pixel's mask are calculated via Eq. \ref{['eq:compactness_of_affinities']}.
• Figure 2: Overall summary of the COCA-Net hierarchy and a single COCA layer.
• Figure 3: Qualitative results of the COCA-Net architecture across six datasets examined in this work. For each dataset, we present four challenging samples along with COCA-Net’s reconstructions. Ground truth segmentation masks, as well as segmentation masks from the decoder and encoder sub-networks, are also included. The mask with the highest intersection with the background segment is shown in gray.
• Figure 4: Visualization of the Non-Overlapping partitioning strategy.
• Figure 5: Qualitative results obtained for real-world datasets. Slot masks predicted by COCA-Net Encoder are shared for Birds (left) and Flowers (right).