Reproduced and visualized results from state-of-the-art models (DRN, LGI, CMINS). Analyzed and compared datasets (ActivityNet-Captions, Charades-STA, TACoS).

Implemented a graph convolutional net to reason the target clip with language syntactics.

Designed an attention mechanism to leverage object detection results to associate frames with nouns in sentences.

Implemented a U-Net image compression model with quantizers and binarizers on skip connections at different levels.

Designed and tuned the training procedure to analyze the effectiveness of each skip connection.

Enabled different trade-offs between compression size and quality by turning on and off some skip connections.