Long-term temporal information yields crucial cues for video action understanding. Previous researches always rely on sequential models such as recurrent networks, memory units, segmental models, self-attention mechanism to integrate the local temporal features for long-term temporal modeling. Recurrent or memory networks record temporal patterns (or relations) by memory units, which are proved to be difficult to capture long-term information in machine translation. Self-attention mechanisms directly aggregate all local information with attention weights which is more straightforward and efficient than the former. However, the attention weights from self-attention ignore the relations between local information and global information which may lead to unreliable attention. To this end, we propose a new attention network architecture, termed as Cascade multi-head ATtention Network (CATNet), which constructs video representations with two-level attentions, namely multi-head local self-attentions and relation based global attentions. Starting from the segment features generated by backbone networks, CATNet first learns multiple attention weights for each segment to capture the importance of local features in a self-attention manner. With the local attention weights, CATNet integrates local features into several global representations, and then learns the second level attention for the global information by a relation manner. Extensive experiments on Kinetics, HMDB51, and UCF101 show that our CATNet boosts the baseline network with a large margin. With only RGB information, we respectively achieve 75.8%, 75.2%, and 96.0% on these three datasets, which are comparable or superior to the state of the arts. •Propose a Cascade multi-head attention Network to construct video representations.•Provide visual analysis for multi-head attention weights.•Achieve performance comparable or superior to the state of the arts.