I'm a postdoc researcher who likes math and computers. The following are very outdated projects from my pre-PhD era (a true pre-LLM time capsule lol):
| Highlights from My Projects | |
|---|---|
| Vision Question Answering (VQA) VQA is a task in computer vision that involves answering questions about an image. My work focused on Bottom-Up and Top-Down Attention Mechanism, optimizing object-level attention. The improved implementation achieved 63.61% accuracy, surpassing the original. |
 |
| Vision Language Navigation (VLN) VLN is a task where agents learn to navigate following natural language instructions. In contrast to VQA, VLN model encodes texts and outputs actions while observing the new information. I explored a robot navigation in multi-floor buildings. |
 |
| Adaptive Zoom Mechanism for Vision Language Navition I integrated an Adaptive Zoom mechanism into VLN, enabling agents to locate large landmarks with wide-FOV vision and identify smaller or distant objects with magnified vision. |
 |
| Trajectory Encoding for Vision Language Navition I designed a model that leverages pre-exploration information in 3D buildings, achieving a 45.8% success rate. (The state of the art was 46.5% at the time.) |
 |
| Simultaneous Localization And Mapping (SLAM) SLAM enables robotic mapping and navigation by constructing an environment map and tracking the robot's location simultaneously. I utilized a particle filter model to create a texture map from a differential-drive robot's two-minute activity in a building. |
 |
| Corner Detection and Sparse Stereo Matching with Epipolar Geometry Employing epipolar geometry, the geometry of stereo vision, I detected and identified corresponding features in images taken from two distinct camera positions. |
 |
| Photometric Stereo for Surface Reconstruction and Phong Illumination for Surface Rendering I used photometric stereo, a computer vision technique, to estimate surface normals under varying lighting conditions. The reconstructed surface was then rendered using the Phong reflection model, a computer graphics method for calculating local illumination. |
 |