- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper presents EventMem, a training-free event-centric memory framework for streaming video understanding with frozen MLLMs. The method consists of three main components. First, Temporal-aware Adaptive K-Means Event Segmentation (TAKC) clusters frame-level features and adaptively determines the number of events. Second, Temporal Segmentation and Aligned Frame Compression (TSAFM) preserves frames with large temporal changes while merging tokens from relatively stable frames. Third, the method maintains a complementary long-term event memory and short-term residual memory, retrieving query-relevant historical events while retaining recent fine-grained visual evidence. The method is evaluated on StreamingBench Real-Time, StreamBench, and OVO-Bench Real-Time using LongVA and Qwen2.5-VL backbones. The reported results show consistent improvements over the corresponding backbones, with particularly noticeable gains on object search and long-term memory tasks. The ablation study also suggests that adaptive event segmentation, token compression, and residual short-term memory contribute complementary improvements.
Critical strengths
-
The long-term event memory plus short-term residual memory design is simple, intuitive, and supported by the ablation study.
-
EventMem is training-free and produces consistent improvements across several streaming-video benchmarks and backbones.
Critical weaknesses
-
The event segmentation mechanism is not sufficiently justified. TAKC determines the number of events from K-means clustering of frame-level visual features, but the number of feature-space clusters does not necessarily correspond to the number of temporal events: a single event may contain multiple visual states, while distinct events may have highly similar frame appearances. This concern is further amplified by the fixed K_{\max}. I am also unclear about the boundary criterion in Eq. (5): maximizing the summed distance to two representative frames does not intuitively correspond to locating the transition boundary. Since event discovery is a central contribution, this needs stronger justification or clarification.
-
The efficiency claim is insufficiently evaluated. The paper motivates EventMem as an efficient streaming-memory framework, but reports almost exclusively task accuracy. It does not provide token counts, compression ratio, memory usage, latency, throughput, or comparisons under matched memory/token budgets. Since EventMem also stores event tokens, key-frame tokens, captions, and residual raw-frame tokens, it is difficult to tell whether the gains come from better memory organization or simply from retaining more information.
Questions for authors
- Could the authors clarify Eq. (5)? Why is the event boundary defined by maximizing the summed distance to two representative frames, and is this exactly what is used in the implementation?
- What justifies treating the silhouette-selected number of feature clusters as the number of temporal events? A single event may contain multiple visual states, while distinct events may have similar frame-level features. It would be useful to know how well the selected K^* correlates with actual event boundaries or event counts.
- Can the authors provide efficiency results under matched memory/token budgets? Retained token counts, compression ratio, memory usage, and latency would help verify the claimed streaming efficiency. The authors should also reconcile the numerical inconsistencies between several tables and the corresponding text.