Hyukjin Kim

• Text-based person retrieval for missing-person and suspect searches where reference images are unavailable and only natural-language descriptions may be provided
• Adding image-grounded attributes generated by a vision-language model produced only modest gains, which led to asking which textual details the model actually uses
• Controlled attribute-ablation experiments on IRRA across CUHK-PEDES, RSTPReID, and ICFG-PEDES: removing upper-body clothing descriptions reduced Rank-1 by 31.6, 25.6, and 43.3 percentage points; the relative importance of garment color and type differed across datasets

• Studied how an action is performed, not just which action: whether local motion patterns capture proficiency differences that global video representations miss
• Contributed core ideas and experimental design for action-guided Prototype Dynamics Margin (PDM) and CORAL-based ordinal calibration
• Combined egocentric and exocentric video with gaze, using atomic-action descriptions as temporal guidance; 54.2% accuracy on Ego-Exo4D, outperforming the strongest evaluated prior baseline (50.1%) by 4.1 percentage points

• Investigated cross-domain generalization conflicts in federated person re-identification
• Developed identity-ranking-based diagnostics in feature space
• Leave-one-client-out experiments to assess individual client contributions to unseen-domain retrieval

• Designed and implemented SwapMixer, a generation-free feature-level augmentation method
• Combines identity representations with pose embeddings from different identities
• Removes the need for a separate pose-conditional image generator

• Baseline: BARON (Aligning Bag of Regions for open-vocabulary object detection)
• Identified a mismatch between teacher and student similarity structures for bags of regions sampled around the same region proposal
• Applied similarity-preserving knowledge distillation to align teacher–student relational representations
• Improved novel-class box AP50 on COCO by +1.2 points (23.1 → 24.3)

• Online clustering with the Sinkhorn–Knopp algorithm
• Multi-view consistency on a ResNet-18 backbone
• Transferred a pretext-trained network to CIFAR-100 classification

• VGG-16 and ResNet — CIFAR-10 classification
• Spatial Transformer Network — distorted MNIST
• FSRCNN — super-resolution
• FCN — semantic segmentation
• RetinaNet with Focal Loss — object detection
• Grad-CAM — AlexNet and VGGNet visualization
• Neural style transfer
• DCGAN — image generation
• CycleGAN — unpaired image translation
• Seq2Seq with attention — machine translation
• Network quantization — QIL with a Straight-Through Estimator
• Self-supervised learning (SwAV) — CIFAR-100

• Lasso and Ridge regression
• Decision Tree and AdaBoost classifiers
• Modular neural networks (forward / backward passes)
• Implemented without ML libraries