TALL: Temporal Activity Localization via Language Query (Supplemental Material)

To directly compare our the temporal regression method with previous state-of-the-art methods on traditional action detection task, we did additional experiments on THUMOS-14. Since THUMOS is a classification task with limited number of action classes, we removed the cross-modal part and trained the localization network with classification loss (cross-entropy loss) and regression loss. We trained a model on the validation set (train set only contains trimmed videos which are not suitable for localization task) and tested it on the test set. The regression model contains 20*2 outputs, corresponding to the 20 categories in the dataset, α is set to 2.0 and 10.0 for non-parameterized and parameterized regression respectively. For each category, we use NMS to eliminate redundant detections in every video, the NMS threshold is set to (tIoU delta), where tIoU = 0.5 and delta=0.2. We report mAP at tIoU=0.5. For training sample generation, we use the same procedure as SCNN [24], we set the high IoU threshold as 0.5 (SCNN used 0.7) and low IoU threshold as 0.1 (SCNN used 0.3) for generating training samples. Note that, our method and SCNN both use C3D features.