The third view with radius 0.3 was further translated to make the points overlapping, representing anoisyview. Then,theentiremodelwas trained with the learning rate of 0.003. We usedβ = 1 for the regularization coefficient. During inference, 100 samples were used to make a prediction. We used the early fusion method to concatenate multi-view features into unimodalfeature.

Inthis work, we extend this line of work to "lazy training" of neural networks - a dominant model in deep learning theory in which neural networks are provably efficiently learnable.

In light of this, we propose an Integration-Decomposition Network (IDN) to implement the above transformations and achievestate-of-the-art performance on widely-used HOI detectionbenchmarks.