Research archive
My papers and projects, with short notes on what each one tries to solve, how it works, and where it falls short.
When more than 95% of local crowd-map blocks are empty, is squared-error regression still the right statistical model?
In briefUse a zero-inflated Poisson likelihood to separate structurally empty regions from discrete counts, without requiring segmentation masks or Gaussian-smoothed targets.
From one second of egocentric video, can an agent identify who will engage, whether the attitude is positive or negative, and what social action comes next?
In briefRepresent face, body, and hand keypoints as graphs, model their temporal evolution, and forecast intent, attitude, and action through a chained task hierarchy.
CLIP was trained to match images and words, not to regress continuous values. Can counting be reformulated so that its recognition knowledge remains useful?
In briefTurn each local count into an unambiguous integer class, match local CLIP image features to natural-language count prompts, and recover a density map from the class probabilities.
How should four heterogeneous video sources be fused, and can the same model remain useful when one or more sources fail?
In briefFuse top/front and depth/infrared features with source-aware self-attention, and randomly mask feature patches during training to simulate missing views or modalities.
If neighbouring in-cabin frames are almost identical, do driver-monitoring systems need 3D video models and every available sensor stream?
In briefUse only the last frame of each 16-frame clip, then measure how top/front views and depth/infrared modalities trade accuracy for latency.
Why build an extra multiscale module if the encoder has already produced features with many receptive-field sizes?
In briefReuse 15 intermediate VGG-16 feature maps, learn spatially varying fusion weights from their contrasts, and decode them with a lightweight saliency-preserving reduction block.
No publications match this tag.