This is an implementation of Tan Yu, et Al’s paper https://www.ijcai.org/proceedings/2017/0635.pdf using Pytorch. The bounding box proposals used for training are generated using an openCV implementation of EdgeBoxes. Mongodb serves as a scalable data store to retain the object proposals indexed by video frames. For each video, we learn a reconstruction model(based on K-Sparse autoencoders) that reconstructs all the proposals. In the test phase, the query image is reconstructed iteratively and the model that reconstructs it the best, would likely contain the object.

Github Link: https://github.com/AshVijay/Instance-retrieval-from-video-sources —