TRECVID 2007 by the Brno Group

1. The runs: • A_brU_1 – features extracted from each frame; SVM per-frame classifier trained on frames in each shot; simple decision tree judging shots based on per-frame results • A_brV_2 – same as A_brU_1, but SVM trained on all training data (the first run divided the training data to training and cross-validation datasets), with SVM configured from the previous run 2. Significant differences between the runs: • As expected, the second run performed generally better, in some cases notably better (which is slightly surprising, because besides the amount of training data, there was no change) 3. Contribution of each component: • The low-level features appear to be good enough, though their number is relatively large and having more time we would experiment with reduction of the feature vector size (now 572 low level features) • We considered using some mid-level features based on existing solutions the group has, such as face detection, car detection, etc., but for time constraints did not employ these in the feature vector • The per-frame classification seems to suffer greatly from mis-annotated frames (whole shots are considered to share the same annotation information in our system) and could be the weakest point of the system • The per-shot decision making seems to be sufficient given the data coming from the per-shot classification 4. Overall comments: • see further in the paper