Research map / Continual learning / Online and replay-based
Live and Learn: Continual Action Clustering with Incremental Views
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes continual action clustering (CAC), which handles incrementally arriving camera views via a category memory library and a consensus partition matrix updated by three-step alternating optimization.
Problem
Multi-view action clustering assumes all camera views are available upfront, which is impractical when views arrive incrementally over time; learning invariant cross-view information in a continual setting is also hard.
Approach
A category memory library B stores cluster centroids learned from historical views, and only a consensus partition matrix is maintained and updated with each new view's basic partition (via late-fusion alignment), avoiding storing or re-accessing all past views; optimized by alternating updates of E, B and W_t.
Key results
- On IXMAS, CAC reaches 68.79% ACC, versus 65.76% for LFMVC and 69.69% for CMVC on WVU (CAC 73.85% ACC on WVU).
- CAC achieves best ACC on 5 of 6 action datasets (e.g., 73.85% on WVU, 49.06% on MSR), outperforming 15 baselines.
- CAC obtains 7.12%, 7.27% and 5.46% ACC gains over L2SC, SCGL and CMVC on IXMAS.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CAC instead of the strongest baseline | IXMAS · IXMAS | ACC: 65.76 -> 68.79 | Helped |
| Use CAC instead of the strongest baseline | IXMAS · IXMAS | NMI: 66.56 -> 65.77 | Hurt |
| Use CAC instead of the strongest baseline | MMI · MMI | ACC: 46.60 -> 46.96 | Helped |
| Use CAC instead of the strongest baseline | MSR · MSR | ACC: 46.88 -> 49.06 | Helped |
| Use CAC instead of the strongest baseline | WVU · WVU | NMI: 66.78 -> 70.53 | Helped |
| Use CAC instead of the strongest baseline | UCLA · UCLA | ACC: 44.95 -> 44.79 | Hurt |
| Use CAC instead of the strongest baseline | UCLA · UCLA | NMI: 44.32 -> 48.52 | Helped |
| Use CAC instead of the strongest baseline | MVTJU · MVTJU | Purity: 75.54 -> 74.18 | Hurt |
| Increase the number of incoming views | CAC · WVU and UCLA | clustering performance | Helped |
| Vary the view arrival order | CAC · WVU and UCLA | clustering performance | No clear effect |
Surprising findings
- The paper reports that view order has little impact on final clustering performance.
Limitations
- Convergence argument is loose (Cauchy-Schwarz bounds only), and CAC is not the fastest method in running time.
Experimental setup
Datasets: IXMAS MMI MSR WVU UCLA MV-TJU
Benchmarks: IXMAS MMI MSR WVU UCLA MV-TJU CCV CiteSeer YouTube
Metrics: ACC NMI Purity running time
Baselines: LFMVC OPMVC Fastmice AGLLFA LKA SMKKM BMVC APMC MIR GCFAgg APADC AONGR
Builds on: late fusion multi-view clustering multi-kernel k-means kernel k-means
Closest papers by meaning
- AdaptCMVC: Robust Adaption to Incremental Views in Continual Multi-view Clustering (CVPR 2025 · similarity 0.86)
- Control-oriented Clustering of Visual Latent Representation (ICLR 2025 · similarity 0.85)
- Coherent Temporal Synthesis for Incremental Action Segmentation (CVPR 2024 · similarity 0.84)
- Unsupervised Action Segmentation via Fast Learning of Semantically Consistent Actoms (AAAI 2024 · similarity 0.84)
- Re-evaluating Continual Learning with Few-Shot Adaptation (NeurIPS 2026 · similarity 0.83)
- Adaptive Visual Scene Understanding: Incremental Scene Graph Generation (NeurIPS 2024 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.