Problem

Given two disjoint camera views, we wish to estimate: (1) their inter-camera correlation, (2) and their spatial-temporal dependencies. Moreover, we aim to answer the question: What visual representations are more effective? Motivation Overcome the unreliability of manually selecting visual features from specific datasets; Explore high-level structural constraints in coding low-level features for associating objects entities (supervised); Employ co-occurrence statistics for constructing more reliable representations (unsupervised). Contributions (1) A systematic investigation into the effectiveness of supervised versus unsupervised feature coding methods for learning inter-camera dependencies; (2) Evaluation of the sensitivity of learning inter-camera time correlation to the size of training data and the quality of scene region decomposition. Methodology (i) Supervised method: Random Forest (RF) [1] for supervised feature coding; (ii) Unsupervised method: Latent Dirichlet Allocation (LDA) [2] for mapping low-level features to code-words that capture topic distributions;

Figure 1: An overview of feature coding comparison for learning inter-camera dependencies. (iii) Time Delayed Dependency Inference: Time Delayed Mutual Information (TDMI) [3] for learning inter-camera dependencies with the aforementioned feature codes; (iv) A new metrics called Mutual Information Margin(MIM) proposed for evaluating different feature coding methods: whereanddenote the TDMI function yielded by the connected and unconnected pairs of regions. Experiments

Figure 2: The layout and example views of an Underground Station (US) dataset (left) and the i-LIDS (right) dataset.

Figure 3: Motion Saliency Maps obtained on the US and the i-LIDS datasets. The selected regions are labelled by black digits.

Table 1: Sensitivity to the length of the training sequence: the average improvement in MIM of different feature coding methods over the k-means vector quantisation based representation. Mean improved MIM (MI-MIM) was computed by averaging individual percentage of improvement over the testing range.

Table 2: Sensitivity to region decomposition: Mean Improved MIM was computed following the same steps as explained in Table 1. Experiment 1: sensitiveness to the size of training data (1) Topic code gave the most favourable results (see Table 1); (2) Suggest that feature coding methods can suppress noisy dependencies between unconnected region pairs. Experiment 2: sensitiveness to the quality of region decomposition (1) Topic code shows the best performance for the US dataset while RF pred for the i-LIDS dataset (see Table 2); (2) Suggest that person count and topic clusters can be useful cues for inter-camera dependency learning. Conclusion: (1) Investigate the effectiveness of supervised (RF) and unsupervised (LDA) feature coding methods for learning inter-camera correlations; (2) RF and LDA coding schemes outperform the k-means vector quantisation in robustness to small training data size; (3) The coded features are more reliable to poor scene region decomposition; (4) Feature coding can suppress noisy dependencies while capture inherent correlations between camera views. References [1] Breiman. Machine Learning, 45(1):5–32, 2001. [2] Blei, Ng, Jordan. J. Machine Learning Research, 3:993–1022, 2003. [3] Loy, Xiang, Gong. IEEE Trans PAMI, 34(9):1799-1813, 2012.