Figure 3.

The pipeline of object-centric anomaly localization framework. After the frame-level anomaly detection stage, we obtain the labels for each frame. Only positive samples (i.e., abnormal frame) are taken into account to localization stage. For each abnormal frame, we run a fast detector to compute bounding boxes (BB) for each object in the sequence. To speed up the computational complexity in practical experiments, we do not re-pass each BB throughout whole framework but we apply directly each BB to error maps. PSNR scores are computed for each one. The BBs yielding minimum values for PSNR scores or values smaller than a threshold refer to the abnormal objects.