Improving semantic video retrieval models by training with a relevance-aware online mining strategy

Alex Falcon, Giuseppe Serra, Oswald Lanz

Abstract: Highlights•Standard contrastive learning (CL) neglects positives outside the training pairs.•CL treats all unpaired elements as irrelevant, even though they are relevant.•We introduce RANP, a general training strategy overcoming these two limitations.•We implement RANP into two novel loss functions, Triplet-RANP and NCE-RANP.•Using RANP, we observe consistent improvements on four methods and four datasets.