From Coarse to Fine-grained Concept based Discrimination for Phrase Detection

Phrase detection requires methods to identify if a phrase is relevant to an image and localize it if applicable. A key challenge in training more discriminative phrase detection models is sampling negatives. However, sampling techniques from prior work focus primarily on hard, often noisy, negatives disregarding the broader distribution of negative samples. To address this problem, we introduce CFCD-Net, a phrase detector that differentiates between phrases through two novels methods. First, we generate groups that consist of semantically similar words we call concepts (eg {dog, cat, horse, ...} vs. car, truck, ...}), and then train our CFCD-Net to discriminate between a region of interest and its unrelated concepts. Second, for phrases containing fine-grained mutually-exclusive words (eg colors), we force the model into selecting only one applicable phrase for each region using our novel fine grained module (FGM). We evaluate our approach on the Flickr30K Entities and RefCOCO+ datasets, where we improve mAP over the state-of-the-art by 1.5-2 points. When considering only the phrases affected by our fine-grained reasoning module, we improve by 3-4 points on both datasets.