In this report, the overview of the runs during the TRECVID 2021 by the ITI-CERTH team are presented. ITI-CERTH participated in the Ad-hoc Video Search (AVS) and Activities in Extended Video (ActEV) tasks. For the AVS task, our participation is based on an attention-based cross-modal deep network architecture. As part of training this architecture, we experimented with a new hard negative mining approach. For the ActEV task, we improve our framework, in terms of more accurate performance, by addressing the classification problem as multi-label rather than a single-label.