School of Engineering Department of Computer Science and Engineering 98 AI meets Big Data: User Analytics and Personalized Recommendation Based on Location Data Supervisor: CHAN Gary Shueng Han / CSE Student: LAI Ming Long / DSCT Course: UROP 1000, Summer Reconstruction of indoor space using sparse-view images has many use cases, including navigation in shopping malls and heritage preservation in museums. In this report, we evaluate the effectiveness of Visual Geometry Grounded Transformer (VGGT) on reconstructing indoor space, particularly shops, using sparseview images. Results indicate that VGGT performs well when there is substantial overlapping in the views. However, when the images have little or no overlap, VGGT cannot accurately resolve the camera angles, and thus make an inaccurate representation of the scene. Furthermore, VGGT cannot generate the unseen areas between the non-overlapping images, affecting the quality of the 3D output. To improve the quality of reconstruction, the next step of the research would be to explore 3D generation methods to generate unseen areas with reasonable quality. AI meets Big Data: User Analytics and Personalized Recommendation Based on Location Data Supervisor: CHAN Gary Shueng Han / CSE Student: LIU Tingsen / DSCT Course: UROP 1100, Summer During this summer’s UROP project, my work mainly focused on AP SSID matching. I used an LLM API (GPT4omini) to achieve semantic matching between SSIDs (Service Set Identifier) and POIs (Points of Interest), thereby roughly obtaining the matching relationship between a given SSID and POI. Based on this, I performed some pre-processing and post-processing operations. In addition, I also conducted manual annotation of the matches in order to quantitatively evaluate the LLM matching results. Moreover, I tried to write a web crawler to obtain the floor plan information of some shops for further research. AI meets Big Data: User Analytics and Personalized Recommendation Based on Location Data Supervisor: CHAN Gary Shueng Han / CSE Student: Mohammad Sufian / COSC Course: UROP 2100, Fall Data collection for indoor localisation tasks can be a time-consuming and labour-intensive task. One technique to overcome this limitation is to generate artificial received signal strength indicator (RSSI) fingerprints from a small initial dataset. The artificial dataset, which captures the underlying distribution of the initial dataset, can be further used to train meaningful models for indoor localisation tasks. In this report, a variational autoencoder (VAE) architecture will be demonstrated, which can provide satisfactory results for generating an artificial WiFi fingerprint dataset for Level 5 of the Hong Kong International Airport based on an initial sample of WiFi fingerprints containing a mixture of both densely and sparsely sampled WiFi fingerprints.
RkJQdWJsaXNoZXIy NDk5Njg=