The whole process of finding landmarks, stabilizing them and annotating the comments with the nearest landmark is now automated.
A couple of issues need to be addressed to ensure correctness of landmark detection. I found them interesting after observations from preliminary testing:
1. The thresholds which determine whether a cluster is a landmark [intra-cluster area closeness + sensor feature closeness], a landmark is a stable one [inter-cluster closeness] have to be learnt over time. Initially they were hard-coded to some value suitable for the conditions under which they were tested.
Solution: To learn them from the clusters got. For example an 80 percentile value of cluster-nearness taken as threshold.
2. Seed Landmark detection module, which the paper argued to be non-essential over a period of time; does in fact affect the initial accuracy of the system. We need to identify these (initial seed) sensor signatures like going up a lift, etc.
Solution: To use a tree like structure to identify the signatures of each place (like in page 6 of the paper).
Implementing the first solution now, should be done with the second one also by Wednesday. Also, I'll be starting testing with my new phone Google Nexus 4 from today. [ This should give us an idea of what these latest smartphones are capable of, with respect to collecting sensor data].
A couple of issues need to be addressed to ensure correctness of landmark detection. I found them interesting after observations from preliminary testing:
1. The thresholds which determine whether a cluster is a landmark [intra-cluster area closeness + sensor feature closeness], a landmark is a stable one [inter-cluster closeness] have to be learnt over time. Initially they were hard-coded to some value suitable for the conditions under which they were tested.
Solution: To learn them from the clusters got. For example an 80 percentile value of cluster-nearness taken as threshold.
2. Seed Landmark detection module, which the paper argued to be non-essential over a period of time; does in fact affect the initial accuracy of the system. We need to identify these (initial seed) sensor signatures like going up a lift, etc.
Solution: To use a tree like structure to identify the signatures of each place (like in page 6 of the paper).
Implementing the first solution now, should be done with the second one also by Wednesday. Also, I'll be starting testing with my new phone Google Nexus 4 from today. [ This should give us an idea of what these latest smartphones are capable of, with respect to collecting sensor data].
No comments:
Post a Comment