We discuss the limitations of the proposed system, along with future research directions. We also highlight a number of areas that require further investigation. Privacy. Although our data collector enables the deletion of collected data, uploading raw images into the server has privacy concerns in practice. Here, one possible solution is to process the classifiers locally in smartphones. For example, instead of uploading raw data, smartphone extracts OCR words, GIST, or SIFT features and uploads these features to server in an unordered manner. To realize this local processing technique, we need an efficient classifier with low complexity and also a job scheduler that would execute classifying/uploading operations when the device is relatively under usage (e.g., night time with battery charging). Meanwhile, one thing we counter-intuitively found in our study is that, many participants are willing to share pictures publically with others. In fact, social network users are already sharing many images, when interacting with others. Therefore, we believe that local data processing may accelerate the active and voluntary participation of end users in crowdsensing framework. Incremental Learning. The proposed system required an non-trivial amount time of data processing in our server (several hours of processing in Intel i7 CPU 860 server with 8GB RAM). The major workload in the server is to process the learning components (i.e., feature extraction and clustering). Considering that we, in our work, used data collected in only one city and the amount of data is rapidly increasing with a widespread use of smartphones and SNS, the data processing issue is practically a big challenge in crowdsensing framework. We believe that an incremental learning scheme would be necessary to handle a large volume of data efficiently in real life. User Participation. The proposed system requires knowledge from crowdsourcing to infer the place name. Our work expands the coverage problem on place naming by applying the SNS knowledge to information collected with the crowdsensing approach. However, the system is not able to learn additional names that are not in SNS. To solve this limitation, the system should somehow support generating new names, or at least provide a method to encourage users to put their knowledge back into SNS. Generating new names is beyond the scope of our research. We, instead, plan to investigate a reward system to induce active user participation in crowdsensing approach. Considering that SNS is missing about 22% of the places visited in daily life (see Section 2.1), an appropriate rewarding mechanism would greatly improve the coverage of place naming.
We discuss the limitations of the proposed system, along with future research directions. We also highlight a number of areas that require further investigation. Privacy. Although our data collector enables the deletion of collected data, uploading raw images into the server has privacy concerns in practice. Here, one possible solution is to process the classifiers locally in smartphones. For example, instead of uploading raw data, smartphone extracts OCR words, GIST, or SIFT features and uploads these features to server in an unordered manner. To realize this local processing technique, we need an efficient classifier with low complexity and also a job scheduler that would execute classifying/uploading operations when the device is relatively under usage (e.g., night time with battery charging). Meanwhile, one thing we counter-intuitively found in our study is that, many participants are willing to share pictures publically with others. In fact, social network users are already sharing many images, when interacting with others. Therefore, we believe that local data processing may accelerate the active and voluntary participation of end users in crowdsensing framework. Incremental Learning. The proposed system required an non-trivial amount time of data processing in our server (several hours of processing in Intel i7 CPU 860 server with 8GB RAM). The major workload in the server is to process the learning components (i.e., feature extraction and clustering). Considering that we, in our work, used data collected in only one city and the amount of data is rapidly increasing with a widespread use of smartphones and SNS, the data processing issue is practically a big challenge in crowdsensing framework. We believe that an incremental learning scheme would be necessary to handle a large volume of data efficiently in real life. User Participation. The proposed system requires knowledge from crowdsourcing to infer the place name. Our work expands the coverage problem on place naming by applying the SNS knowledge to information collected with the crowdsensing approach. However, the system is not able to learn additional names that are not in SNS. To solve this limitation, the system should somehow support generating new names, or at least provide a method to encourage users to put their knowledge back into SNS. Generating new names is beyond the scope of our research. We, instead, plan to investigate a reward system to induce active user participation in crowdsensing approach. Considering that SNS is missing about 22% of the places visited in daily life (see Section 2.1), an appropriate rewarding mechanism would greatly improve the coverage of place naming.
การแปล กรุณารอสักครู่..
