inflated 3d convnet

The Inflated 3D ConvNet (I3D, Carreira and Zisserman, 2017) solved the central pain of 3D CNNs, that they are data-hungry and lack good pretraining, with one elegant trick: take a powerful 2D image network already trained on ImageNet and 'inflate' every 2D filter into 3D by replicating it across the time axis. A 2D k×k filter becomes a 3D k×k×k filter whose k temporal slices are copies of the original, divided by k so the responses stay calibrated. The new 3D network thus inherits all the visual knowledge the 2D network learned from millions of labeled images, instead of starting from random weights.

The normalization is the clever part and gives a precise invariant. Because each inflated filter averages its (identical) copies over time and is scaled by 1/k, feeding a 'boring' video, one where every frame is an identical still image, makes the inflated 3D network produce exactly the same activations, and the same prediction, as the original 2D network applied to that single frame. So before any video training, I3D already behaves like a competent image model; fine-tuning on video (Carreira and Zisserman trained on the then-new Kinetics dataset) only has to teach it the additional temporal structure. The best result combined two inflated streams (RGB and optical flow), as in two-stream I3D.

I3D's impact was to establish a standard recipe that the field still follows: inflate a strong 2D backbone, pretrain or co-train on Kinetics, then transfer to smaller action datasets. This bootstrapping made 3D models both practical and state-of-the-art, and its descendants (SlowFast, X3D) and the later shift to video transformers all build on the same lineage of ImageNet/Kinetics pretraining that I3D made mainstream.