We introduce optical Flow transFormer, dubbed as FlowFormer, a transformer-based neural
network architecture for learning optical flow. FlowFormer tokenizes the 4D cost volume built
from an image pair, encodes the cost tokens into a cost memory with alternate-group
transformer (AGT) layers in a novel latent space, and decodes the cost memory via a
recurrent transformer decoder with dynamic positional cost queries. On the Sintel
benchmark, FlowFormer achieves 1.144 and 2.183 average end-ponit-error (AEPE) on the …

As shown in Table 1, we conduct additional ablation experiments on the alternative-group
transformer (AGT) layer. For intra-cost-map aggregation layer, since the number and
dimension of latent cost tokens are fixed, we test on replacing our design with MLP-Mixer
[4](2nd row), which is a state-of-the-art MLP-based architecture. We also substitute
ConvNeXt [3] for transformer in inter-cost-map aggregation (3rd row). Furthermore, we
replace both transformers with MLP and ConvNext (4th row). Replacing transformer layers …