Video analysis relies on neural networks that process both spatial and temporal information. Three widely used architectures are 3D Convolutional Neural Networks (3D CNNs), Two-Stream Networks, and Transformer-based models. Each addresses the challenge of capturing motion and context across video frames, though they differ in design and computational demands. Hybrid approaches combining these architectures or integrating recurrent components are also common, depending on the task and available resources.
3D CNNs extend traditional 2D CNNs by adding a temporal dimension to convolutions, enabling direct spatiotemporal feature learning. For example, the C3D model applies 3x3x3 kernels across frames to extract motion patterns, making it effective for action recognition in short clips. A more advanced variant, I3D (Inflated 3D CNN), “inflates” 2D convolutional kernels pretrained on ImageNet into 3D, improving performance on datasets like Kinetics. While 3D CNNs capture local motion well, they are computationally heavy due to processing multiple frames simultaneously. To mitigate this, techniques like Pseudo-3D (P3D) or R(2+1)D separate spatial and temporal convolutions, reducing parameters while maintaining accuracy.
Two-Stream Networks process spatial (RGB frames) and temporal (motion) data in parallel. The original Two-Stream architecture by Simonyan and Zisserman uses a spatial stream for scene analysis and a temporal stream fed with optical flow (pixel motion vectors) to capture movement. Modern variants like SlowFast Networks optimize this by splitting processing into two pathways: a “slow” branch for high-resolution spatial features and a “fast” branch for low-resolution motion cues. This design balances efficiency and accuracy, leveraging pretrained 2D CNNs for spatial features. TSN (Temporal Segment Networks) further improves temporal modeling by sampling sparse frame segments and aggregating predictions, reducing computation while maintaining context.
Transformer-based models like TimeSformer and Video Swin Transformer adapt self-attention mechanisms for video. TimeSformer divides video into spacetime patches and applies attention across both dimensions, enabling long-range dependency modeling without convolutions. Video Swin Transformer uses a hierarchical approach with shifted windows to reduce computational costs. These models excel in tasks requiring global context, such as long-form activity recognition. For sequence modeling, hybrid architectures like CNN-LSTM combine convolutional layers (for per-frame features) with recurrent layers (e.g., LSTM) to track temporal dynamics, useful for tasks like gesture prediction. Developers often prioritize 3D CNNs or Transformers for accuracy but may opt for Two-Stream or hybrid models when balancing speed and resource constraints.