We propose Extended Co-occurrence HOG (ECoHOG) feature and integrate it with dense sampling and dense feature extraction approach in order to improve accuracy of fine-grained activity recognition. The co-occurrence feature clearly extracts an object's shape by focusing on co-occurrence of image gradients at the pairs of image pixels and in that way reduces false positives. We extend this feature by adding sum of the magnitude of the gradients as co-occurrence elements. This results in giving the importance to the object boundaries and straightening the difference between the moving foreground and static background. In addition we apply this descriptor on the dense trajectories and test it for fine grained activity recognition.