Transition frequency between cameras in the WNMF database.

Contexts in source publication

Context 1

... successful MCTF model can address this issue by preempting the location of an object-of-interest in a distributed camera network, thereby enabling the system to monitor only selected cameras intelligently. We envision an MCTF model to be an additional component of a full multi-camera monitoring system, complementing the existing methods for de- Figure 3. Annotation method. ...

Context 2

... manual annotation, we propose a semi-automated method that uses a combination of off-the-shelf models for detection, tracking, and person RE-ID. These results are then manually verified to ensure that proposed tracks are accurate and correct cross-camera correspondences for pedestrians are found. An overview of this annotation method is shown in Fig. 3, which consists of the following three ...

Context 3

... within a manually specified time-window γ to cut down the search space of possible matches, i.e., we compare only those tracklets which satisfy t E − t D < γ. As we set γ = 12 seconds, the matches are generally from neighboring cameras in the network. We confirmed this by comparing the camera transitions with respect to the network topology in Fig. 3. Our annotation method results in a set of cross-camera transitions T = {(E t , D t ...

Context 4

... frequent transition. Using the transition frequency matrix computed earlier (Fig. 3), we predict the next camera as the most frequent next camera of observation from its corresponding ...

Context 5

... of 20% between fully-connected layers. Discussion. Table 2 shows next camera prediction results. Predicting the correct camera in the top 3 is a straightforward problem in our dataset, given the structured camera setup and junctions with at most 3 exits. Predicting the most frequent transition using the transition matrix from the training data (Fig. 3) attains modest performance, although learned methods perform better, particularly in terms of top-1 accuracy. We suspect this is due to the past trajectory information in one camera view being informative of the person's future trajectory in a way that is not captured by other baselines. We observe moderate improvement in using ...