My research
PhD at Nanyang Technological University, supervised by Assoc. Prof. Deepu Rajan. Thesis: "Learning Compositionality for Image Classification and Segmentation" — three contributions, all aimed at cutting how much a human has to annotate.
- Contribution 01
Multimodal few-shot classification without attribute embedding
Uses attributes directly rather than embedding them, so interpretability survives. Published in the EURASIP Journal on Image and Video Processing, 2024.
- Contribution 02
Part segmentation from object masks and class labels
Drops pixel-level part annotation entirely, using superpixels and a weakly-supervised part classifier over a ResNet-50 detector.
- Contribution 03
Part segmentation from class labels alone
Removes the object masks too, grouping superpixels with DINO-ViT features. Groundwork toward annotation-free part segmentation.
