The Visual Bottleneck: Sparse-Frame Adaptation of MLLMs for Joint Spatial-Temporal Video Grounding
2026-07-27 • Computer Vision and Pattern Recognition
Computer Vision and Pattern RecognitionArtificial Intelligence
AI summaryⓘ
The authors studied how to improve video moderation systems that have to quickly find policy violations using very few frames from each video, unlike training which uses many frames. They found that the main problem is how video features are extracted when only a few frames are available. By fine-tuning just a small part of the visual model, they greatly improved accuracy, better than using large language models without training. They also showed that selecting frames near event boundaries helps a lot. Overall, their work suggests focusing on training strategies is more important than just using bigger models when dealing with sparse video frames.
video moderationtemporal localizationmultimodal large language models (MLLMs)visual feature extractionVision Transformer (ViT)temporal mean Intersection over Union (mIoU)frame samplingfine-tuningzero-shot learningsparse frame input
Authors
Jiameng Zhang, Srikanth Madikeri
Abstract
Large-scale video platforms process millions of uploads hourly, requiring moderation systems that can localize when and where policy violations occur within each video. Processing every frame is infeasible at scale, so systems are constrained to sparse inputs of 8 to 16 frames per video. Yet state-of-the-art multimodal large language models (MLLMs) are pretrained on dense sequences of hundreds of frames, creating a fundamental mismatch between training and deployment conditions. This mismatch causes severe performance collapse: the Qwen3-VL 8B model drops from 56.0% to 22.3% temporal mIoU when frames are reduced to 16, a 60.2% relative degradation. We present a systematic empirical study of training strategies to close this gap for spatial-temporal video grounding. Our results suggest that visual feature extraction is the dominant bottleneck under sparse-frame inputs. Adapting only the final three ViT layers, 4% of total parameters, achieves 68.8% temporal mIoU and surpasses a zero-shot 8B model using dense inputs by 12.8 points. Language model fine-tuning, by contrast, offers negligible or negative returns. A boundary-aware sampling strategy, Hybrid16, further improves temporal mIoU by 26 points over uniform sampling when temporal boundaries are available. We conclude that for sparse-frame video grounding, training strategy dominates model scale: a fine-tuned 2B model consistently outperforms a zero-shot 8B model, with or without dense frame access.