Temporal Moment Localization is a challenging multi-modal task which aims to identify the start and end timestamps of a moment of interest in an input untrimmed video, given a query in natural language. Solving this task correctly requires understanding the temporal relationships in the entire input video, but processing such long inputs and reasoning about them is memory and computationally expensive. In light of this issue, we propose Stochastic Bucket-wise Feature Sampling (SBFS), a stochastic sampling module that allows methods to process long videos at a constant memory footprint. We further combine SBFS with a new consistency loss to propose Locformer, a Transformer-based model that can process videos as long as 18 minutes. We test our proposals on relevant benchmark datasets, showing that not only can Locformer achieve excellent results, but also that our sampling is more effective than competing counterparts. Concretely, SBFS consistently improves the performance of prior work, by up to 3.13% in the mean temporal IoU, leading to a new state-of-the-art performance on Charades-STA and YouCookII, while also obtaining up to 12.8x speed-up at testing time and reducing memory requirements by up to 5x.

Existing image captioning models do not generalize well to out-of-domain images containing novel scenes or objects. This limitation severely hinders the use of these models in real world applications dealing with images in the wild. We address this problem using a flexible approach that enables existing deep captioning architectures to take advantage of image taggers at test time, without re-training. Our method uses constrained beam search to force the inclusion of selected tag words in the output, and fixed, pretrained word embeddings to facilitate vocabulary expansion to previously unseen tag words. Using this approach we achieve state of the art results for out-of-domain captioning on MSCOCO (and improved results for in-domain captioning). Perhaps surprisingly, our results significantly outperform approaches that incorporate the same tag predictions into the learning algorithm. We also show that we can significantly improve the quality of generated ImageNet captions by leveraging ground-truth labels.