Dense Video Captioning with Convolution-Group Structure

Dense video captioning aims to accurately localize every event within a video and generate natural language captions for each. In recent years, numerous Transformerbased DVC approaches have demonstrated promising performance. However, due to the inherent complexity of DVC datasets—which often include rich content, diverse scenes, dynamic background variations, and frequent shifts in human behavior—models must exhibit stronger capabilities in temporal modeling, semantic understanding, and language generation. Some previous methods have shown limited effectiveness in handling these highly complex video scenarios, potentially due to insufficient supervision during training. To address this limitation, we propose Dense Video Captioning with Convolution-Group Structure (CoG). In the decoding stage, CoG employs a one-to-many group decoding framework to enhance the model's expressive power in challenging video datasets. Furthermore, in the feature encoding stage, we adopt a fully convolutional Scalable-Granularity Perception (SGP) feature encoder, aiming to improve the accuracy of temporal event localization. We evaluate our approach on two benchmark datasets, ActivityNet Captions and YouCook2. Experimental results demonstrate that CoG achieves notable improvements across BLEU4, METEOR, and CIDEr.