(From left) Lee Sun-ho, master's student; Professor Cho Sung-in

The research paper by Lee Sun-ho and Cho Sung-in, "Remembering Across Blocks: Topology-Conditioned Block-Progressive Memory for Skeleton-Based Action Recognition," has been accepted as a regular paper at the European Conference on Computer Vision (ECCV) 2026, the most prestigious international conference in the field of computer vision.

This paper addresses skeleton-based action recognition, the problem of recognizing actions from sequences of human joint coordinates. Skeleton-based action recognition is an important area of computer vision research that requires analysing body structure and temporal motion information together, and graph convolutional network (GCN) based methods have recently been studied widely.

Existing GCN-based methods treat joints as nodes of a graph and learn long-range dependencies by repeatedly performing one-hop message passing between adjacent joints. They also refine representations progressively through block-wise learned topology across multiple GCN blocks. However, repeated graph aggregation can cause an over-smoothing problem in which node features become increasingly similar in deeper blocks. In addition, fine-grained motion cues formed in early blocks may be diluted during subsequent aggregation and thus not sufficiently reflected in the final embedding.

To address these problems, the research team proposes a topology-conditioned block-progressive memory. The proposed method stores the complementary representation formed at each block along the GCN block axis in memory and fuses the accumulated memory state into the final embedding, thereby mitigating the representation degradation caused by repeated graph aggregation.

Figure 1. Overall architecture of the proposed method. (a) Baseline structure based on a conventional GCN feature extractor. (b) The proposed topology-conditioned block-progressive memory module.

The proposed memory module maintains a unique memory state for each sample and updates the per-block memory by minimizing a topology-conditioned reconstruction loss. The reconstruction error measures how well the previous memory explains the current block representation while simultaneously capturing newly emerging discriminative cues. This allows the model to draw on the fine-grained joint representations and structural information formed across multiple blocks rather than relying on the representation of the final block alone.

The paper also supports two modes of operation at inference. Memory-free mode uses the memory module only during training and performs inference at test time without additional memory updates, while memory-enabled mode applies the same memory update rule at test time as well to obtain further performance gains.

In experiments, the proposed method achieved state-of-the-art performance on large-scale skeleton-based action recognition benchmarks including NTU RGB+D, NTU RGB+D 120 and Kinetics-Skeleton. This demonstrates that remembering the useful representations formed at each GCN block and exploiting them for final recognition is effective in improving skeleton-based action recognition performance.

ECCV, together with CVPR and ICCV, is one of the foremost international conferences representing the field of computer vision, and is a prestigious venue at which the latest research results across computer vision are presented. ECCV 2026 will be held in Malmö, Sweden, from 8 to 12 September 2026.

 

[References]

Paper title: Remembering Across Blocks: Topology-Conditioned Block-Progressive Memory for Skeleton-Based Action Recognition

Authors: Lee Sun-ho (first author, Sogang University), Han Sang-hoon (co-author, Sogang University), Nam Hyuk (co-author, Sogang University), Cho Sung-in (corresponding author, Sogang University)

The European Conference on Computer Vision (ECCV) 2026 website: https://eccv.ecva.net/

Lab homepage: https://sites.google.com/view/csi2267svm/