Publications

3D human motion correction via learning from physically simulated data (working title)

Colin, M., et al. • Under preparation — CVPR 2027 submission • 2026

First author. 3D human motion reconstructed from monocular video exhibits systematic artifacts: penetration, floating, foot skating. I train a correction model on a paired clean ↔ degraded corpus that I build by rendering and then re-estimating motion, in which physics is used solely as training data — neither in the loss function nor in the architecture. Work carried out at the University of Alberta's Vision & Learning Lab (Prof. Li Cheng).

Physically plausible multi-person 3D motion generation (working title)

Wang, Y., Colin, M., et al. • Under preparation — ICLR 2027 submission • 2026

Co-author on ongoing research led by Yilin Wang at the University of Alberta's Vision & Learning Lab (Prof. Li Cheng). Music- and text-prompt-conditioned multi-person 3D dance generation; my contribution focuses on the physical plausibility of the motion (ground-plane estimation, contacts, foot-skating, interpenetration, physics-based metrics on AIOZ-GDANCE).

Performances and Explainability of ViT and CNN Architectures: An Empirical Study Using LIME, SHAP, and GradCam

Colin, M., Chraibi Kaadoud, I. • RJCIA 2024 (PFIA), La Rochelle • 2024

In recent years, explainable AI has been presented as the main solution for building trust between users and AI systems. To investigate this hypothesis, we propose an empirical study on the link between the performance and explainability of four computer vision algorithms: ViT, ResNet50, VGG16 and InceptionV3. Our study uses three local explainability methods: LIME, SHAP and GradCam. We show that, while explainable AI can be a tool for challenging the artificial representation of an algorithm and its behavior, it can also present robustness problems or contradictory information that can undermine trust. Our results show that by combining several explainability methods to explain a single prediction, it is possible to verify the reliability of the explanations and the information extracted.

My contribution: experimental protocol design, implementation of the models and explainability methods, analysis of the results, and writing. Written in my own time, teaching myself the scientific method.

@inproceedings{colin2024performances,
  title     = {Performances et explicabilit{\'e} de ViT et d'architectures CNN : une {\'e}tude empirique utilisant LIME, SHAP et GradCam},
  author    = {Colin, M{\'e}lissa and Chraibi Kaadoud, Ikram},
  booktitle = {RJCIA 2024, Plateforme Fran{\c{c}}aise pour l'Intelligence Artificielle (PFIA)},
  year      = {2024},
  address   = {La Rochelle, France},
  url       = {https://hal.science/hal-04641791v1}
}