Agentic Video Summarization via Self-Reflecting Multimodal Understanding
A multimodal agent workflow summarizes videos through autonomous verification and reflection.
Open research
Research implementations, datasets, and benchmarks, with links to the work behind them.
More on GitHub .
A multimodal agent workflow summarizes videos through autonomous verification and reflection.
Decoupled identity and human-parsing features improve person retrieval under occlusion.
Identity shifting protects personal appearance while preserving identity relationships needed for person re-identification.
Hierarchical representations recognize objects directly in vector graphics, supported by a chart understanding benchmark.
Dynamic invisible triggers reveal backdoor vulnerabilities when person re-identification models encounter previously unseen identities.
An energy-aware benchmark studies trade-offs between model quality and the cost of architecture and hyperparameter search.
Weakly supervised human co-parsing aligns visible body regions to improve occluded person re-identification.
A label-aware detector and the CLCXray dataset address overlapping objects in X-ray security imagery.
Incrementally generated occlusions train person re-identification models to suppress occlusion-induced distractions.
Recurrent feedback alternately updates spectral and spatial features for hyperspectral image classification.
A densely connected spectral–spatial convolutional network improves training efficiency and hyperspectral classification accuracy.