I3D

I3D, or Inflated 3D ConvNet, is a convolutional neural network architecture designed for video action recognition. It was introduced by Joao Carreira and Andrew Zisserman in 2017 as an extension of 2D convolutional networks into the spatiotemporal domain by inflating 2D kernels into 3D. The key idea is to take established 2D architectures, such as Inception networks, inflate the spatial kernels to include a temporal dimension, and initialize the weights from pretraining on image data. This approach allows the model to learn both spatial and temporal features from video while leveraging large-scale image datasets.

The architecture uses standard Inception modules adapted to 3D convolutions, featuring multiple branches that capture features at different scales, followed by pooling and a final classification layer. I3D can operate on RGB frames, and a two-stream variant combines RGB with optical flow to capture motion information. In practice, I3D models are often pretrained on large video datasets such as Kinetics and then fine-tuned on target tasks.

I3D achieved strong results on benchmark video datasets and became a common baseline for action recognition research. It demonstrated the effectiveness of inflating 2D networks to 3D for video understanding and influenced subsequent work in efficient video models, including variants that factorize 3D convolutions or fuse temporal information in alternative ways. While computationally intensive, I3D remains a reference point in the study of spatiotemporal feature learning and continues to inform research in video analysis and related fields.