Video Feature Global Pooling - Facebook

Facebook has developed a video copy detection system that is capable of generating feature vectors for fixed-length video clips. However, using clip-level features typically entails comparing all the different possible pairs of clips, a task that gets expensive as the length of the video increases.

I worked on the Facebook AI Integrity Team to pool an unknown-length list of clip-level video features into one final fixed-dimension feature vector that represented all the content in the video. This global feature vector could then be used to efficiently perform copy detection. During my project, I investigated the performance of various generic pooling methodologies, and afterward designed specialized pooling strategies that leveraged the structure of the copy detection problem to improve pooling performance beyond that of the generic strategies.

• Completed on: June-August 2021
• Skills: PyTorch/Pandas/Distributed Computing
• Company: Facebook