Research map / Video / Action recognition
Intentional Evolutionary Learning for Untrimmed Videos with Long Tail Distribution
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 an ICCA loss for long-tail, concept-drifting untrimmed video streams and an intention-oriented evolutionary learning model that predicts current action, next action, and remaining duration on THUMOS14 and ActivityNET v1.3.
Problem
Predicting human intentions and their evolution (from which action to which action, and when) in untrimmed videos, which have long-tail class distributions and concept drift across video streams.
Approach
A class-reweighted cross-entropy loss (ICCA) that uses instance-level confidence from a newer data stream and per-class accuracy from the previous epoch to adapt weights to distribution drift, combined with an RNN-based intention evolution model that outputs next action, evolution time, and an intention-coherence constraint.
Key results
- On THUMOS14 (untrimmed), ours reaches 67.9 top-1 recognition and 64.8 top-1 prediction, versus 60.1/56.8 for latent goal and 59.2 top-5 for RU-reg-style baseline comparisons.
- On ActivityNET v1.3 (untrimmed), ours reaches 65.0 top-1 recognition and 34.3 top-1 prediction with time accuracy 27.04, versus 14.50 for latent goal.
- Using ICCA loss on THUMOS14 (trimmed setting) raises mean precision for prediction to 54.61 versus 50.10 for CDB loss.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace CDB loss with ICCA loss | the proposed method · THUMOS14 | top-1 accuracy, recognition: 66.1 -> 67.3 | Helped |
| replace CDB loss with ICCA loss | the proposed method · THUMOS14 | top-1 accuracy, prediction: 65.8 -> 67.3 | Helped |
| replace weighted-CE with ICCA loss | the proposed method · ActivityNET v1.3 | top-1 accuracy, recognition: 63.9 -> 65.0 | Helped |
| replace weighted-CE with ICCA loss | the proposed method · ActivityNET v1.3 | top-1 accuracy, prediction: 36.1 -> 34.3 | Hurt |
| use the proposed method rather than Latent goal | untrimmed video intention estimation · THUMOS14 | top-1 accuracy, recognition: 60.1 -> 67.9 | Helped |
| use the proposed method rather than Latent goal | untrimmed video intention estimation · THUMOS14 | top-1 accuracy, prediction: 56.8 -> 64.8 | Helped |
| use the proposed method rather than Latent goal | untrimmed video intention estimation · ActivityNET v1.3 | time accuracy: 14.50 -> 27.04 | Helped |
| set ICCA fusion frequency to 5 rather than 10 (ablation) | ICCA loss · THUMOS14 | top-1 accuracy, recognition: 67.3 -> 66.3 | Hurt |
| set ICCA fusion frequency to 20 rather than 10 (ablation) | ICCA loss · THUMOS14 | top-1 accuracy, recognition: 67.3 -> 64.7 | Hurt |
| set ICCA fusion frequency to 30 rather than 10 (ablation) | ICCA loss · THUMOS14 | top-1 accuracy, recognition: 67.3 -> 63.7 | Hurt |
| use mean rather than max for the ICCA weight strategy (ablation) | ICCA loss · THUMOS14 | top-1 accuracy, recognition: 67.3 -> 57.9 | Hurt |
| use sum rather than max for the ICCA weight strategy (ablation) | ICCA loss · THUMOS14 | top-1 accuracy, recognition: 67.3 -> 66.0 | Hurt |
Surprising findings
- Focal loss and other standard long-tail losses sometimes outperform on some metrics, and the best mean precision depends on the choice of error function and weight strategy.
Limitations
- Time-of-evolution estimation accuracy is still limited (time accuracy about 38% on THUMOS14 and 27% on ActivityNET).
- Evaluated only on two datasets using pre-extracted I3D features and a fixed sliding window setup.
Experimental setup
Datasets: THUMOS14 ActivityNET v1.3
Benchmarks: THUMOS14 ActivityNET v1.3
Metrics: top-1 accuracy top-5 accuracy mean precision time accuracy
Baselines: focal loss weighted-CE EQLv1 EQLv2 CDB loss RULSTM latent goal RU-reg
Builds on: CDB loss EQLv2 RULSTM Latent goal I3D
Closest papers by meaning
- AntGPT: Can Large Language Models Help Long-term Action Anticipation from Videos? (ICLR 2024 · similarity 0.83)
- WorldReasonBench: Human-Aligned Stress Testing of Video Generators as Future World-State Predictors (NeurIPS 2026 · similarity 0.83)
- Can't Make an Omelette Without Breaking Some Eggs: Plausible Action Anticipation Using Large Video-Language Models (CVPR 2024 · similarity 0.82)
- Videos are Sample-Efficient Supervisions: Behavior Cloning from Videos via Latent Representations (NeurIPS 2025 · similarity 0.82)
- EvoGround: Self-Evolving Video Agents for Video Temporal Grounding (NeurIPS 2026 · similarity 0.82)
- Video In-context Learning: Autoregressive Transformers are Zero-Shot Video Imitators (ICLR 2025 · similarity 0.82)
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.