Hieu Le

Hieu Le (Minh Hiếu Lê)

Computer Vision Lab - EPFL
BC300, Faculté Informatique et Communications – IC
Ecole Polytechnique Fédérale de Lausanne – EPFL

Email: hle@cs.stonybrook.edu
[Google Scholar] [Github]

I am a postdoc at CVLab EPFL working with Prof. Pascal Fua and Mathieu Salzmann. Before that, I obtained my Ph.D. in computer science from Stony Brook Univeristy working with Prof. Dimitris Samaras and spent two years as an applied scientist at Amazon Robotics - Boston.

My research focuses on building reliable and controllable machine vision systems, combining deep learning with structure, uncertainty, and interpretability to develop models that are not only high-performing, but also transparent, trustworthy, and adaptable in real-world applications.

Structure-guided learning: I incorporate physical or biological constraints and semantic structure into neural networks to improve robustness and control. This includes work on physics-based image enhancement, and controllable generation/modification of image and 3D shapes.
Uncertainty estimation: I design efficient techniques to measure model confidence without relying on ensembles. From iterative refinement networks to auxiliary classifiers, my methods enhance reliability in tasks like image classification, segmentation, and 3D shape optimization.
Explainability in generative models: I explore how latent spaces relate to sample quality and diversity, enabling better control over outputs in models like Stable Diffusion. This helps make generative systems more predictable, interpretable, and useful.

News

[2025] A paper on personalized scanpath prediction has been accepted to CVPR 2025.

[2025] A paper on identifying reliable seeds for diffusion models has been accepted to ICLR 2025 as a Spotlight paper.

[2024] Two papers accepted to WACV 2025: Material-Consistent Shadow Edges, and Multi-Class Object Counting.

[2024] Pseudo-Density has been accepted to TMLR. Similar to Latent Density Score, this paper reveals the learning mechanism of generative models.

[2024] Four papers accepted to ECCV 2024: Latent Density Score, Neural Surface Localization, Pseudo-Pixel Weighting, and Beyond Pixels.

[2024] A paper on 3D heart generation has been accepted to MICCAI 2024.

[2024] Our paper on uncertainty estimation has been accepted to ICML 2024.

[2024] Zigzag has been accepted by TMLR. Zigzag is a universal uncertainty estimation method that requires only two inference passes.

[2023] Two papers accepted to CVPR 2023: Zero-Shot Object Counting and Few-Shot Object Detection.

[2023] Selected as an outstanding reviewer at ACCV 2022.

[2022] Our paper on Few-Shot Classification via Generating Representative Prototypes has been accepted to CVPR 2022.

[2022] Our journal paper on Physics-Based Shadow Removal has been accepted by TPAMI 2022.

[2022] Our journal paper on detecting penguin colonies from satellite images has been accepted by Remote Sensing 2022.

[2021] Our paper on Few-Shot Classification has been accepted to ICCV 2021.

[2021] Outstanding Reviewer - CVPR 2021.

[2020] Ph.D. completed.

[2020] Outstanding Reviewer - ECCV 2020.

[2020] Our paper on Weakly-Supervised Shadow Removal has been accepted to ECCV 2020. Project Page.

[2019] Our paper on Shadow Removal has been accepted to ICCV 2019. Project Page.

[2019] Our paper on Semi-Supervised Segmentation has been accepted to CVPR 2019 - CV4GC.

[2019] Our paper on Policy Mining Using Neural Networks has been accepted to SACMAT 2019.

[2018] Two papers accepted to ECCV 2018: ADNet for shadow detection and Iterative Crowd Counting.

[2017] Our paper on Object Co-Localization has been accepted to ICCV 2017, Workshop on CEFRL, Venice, Italy.

[2017] Our poster on Object Co-Localization was presented at VSS 2017 (Florida, US). [Poster]

[2016] Geodesic Features for Video Segmentation has been accepted to ACCV 2016.

Publications

	Few-shot Personalized Scanpath Prediction Ruoyu Xue, Jingyi Xu, Sounak Mondal, Hieu Le, Greg Zelinsky, Minh Hoai, Dimitris Samaras, 2025 IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2025. [Preprint]
	Enhancing Compositional Text-to-Image Generation with Reliable Random Seeds Shuangqi Li, Hieu Le, Jingyi Xu, Mathieu Salzmann, 2025 International Conference on Learning Representations (ICLR), 2025. [Preprint] TL;DR: Noises are important for diffusion-based models. We show that some random seeds are much more reliable than others, which can be used to generate useful training data.
	Learning to Count from Pseudo-Labeled Segmentation Jingyi Xu, Hieu Le, Dimitris Samaras Winter Conference on Applications of Computer Vision (WACV), 2025. [Preprint] Existing methods suffer from the counting-everything issues. We introduce a benchmark with multiple countable objects in each image and show that we can mitigate this issues by using synthetic data.
	Shadow Removal Refinement via Material-Consistent Shadow Edges Shilin Hu, Hieu Le, ShahRukh Athar, Sagnik Das, Dimitris Samaras Winter Conference on Applications of Computer Vision (WACV), 2025. [Preprint][ Project Page] [ Testing Set ] We proposed a new shadow removal method leveraging supervision from the shadow edges. Plus, we introduced a new benchmark for shadow removal with no shadow-free images needed!
	Assessing Sample Quality via the Latent Space of Generative Models Jingyi Xu, Hieu Le, Dimitris Samaras European Conference on Computer Vision (ECCV), 2024. [ Preprint] [ Code] [ Project Page] Latent points located in dense regions of the latent manifold tend to produce higher-quality samples, while those in sparser regions yield lower-quality ones.
	Controlling the Fidelity and Diversity of Deep Generative Models via Pseudo Density Shuangqi Li, Chen Liu, Tong Zhang, Hieu Le, Sabine Susstrunk, Mathieu Salzmann European Conference on Computer Vision (TMLR), 2024. [ Openreview]
	Neural Surface Localization for Unsigned Distance Fields Federico Stella, Nicolas Talabot, Hieu Le, Pascal Fua European Conference on Computer Vision (ECCV), 2024. [ Preprint] [ Project Page]
	Weighting Pseudo-Labels via High-Activation Feature Index Similarity and Object Detection for Semi-Supervised Segmentation Prantik Howlader, Hieu Le, Dimitris Samaras European Conference on Computer Vision (ECCV), 2024. [ Preprint] [ Code]
	Beyond Pixels: Semi-Supervised Semantic Segmentation with a Multi-scale Patch-based Multi-Label Classifier Prantik Howlader, Srijan Das, Hieu Le, Dimitris Samaras European Conference on Computer Vision (ECCV), 2024. [ Preprint] [ Code]
	Generating Anatomically Accurate Heart Structures via Neural Implicit Fields Jiancheng Yang, Ekaterina Sedykh, Jason Adhinarta, Hieu Le, Pascal Fua MICCAI 2024 [Paper] [Poster]
	Enabling Uncertainty Estimation in Iterative Neural Networks Nikita Durasov, Doruk Oner, Jonathan Donier, Hieu Le, Pascal Fua International Conference on Machine Learning (ICML), 2024. [ Preprint] [ Code]
	Zigzag: Universal Sampling-free Uncertainty Estimation Through Two-Step Inference Nikita Durasov,Nik Dorndorf, Hieu Le, Pascal Fua Transactions on Machine Learning Research (TMLR), 2024. [ Openreview] [ Code]
	Zero-Shot Object Counting Jingyi Xu, Hieu Le, Vu Nguyen, Viresh Ranjan, Dimitris Samaras IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. [ Code] [ Stable-Diffusion extension]
	Generating Features With Increased Crop-related Diversity For Few-shot Object Detection Jingyi Xu, Hieu Le, Dimitris Samaras IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2023. [ Code (TBD)]
	Generating Representative Samples for Few-Shot Classification Jingyi Xu, Hieu Le IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2022. [Code] [Paper ] [ Blog post ]
	Physics-based Shadow Image Decomposition for Shadow Removal Hieu Le, Dimitris Samaras. IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2022. [Github ] [Paper ]
	A convolutional neural network architecture designed for the automated survey of seabird colonies Hieu Le, Dimitris Samaras, Heather J. Lynch Remote Sensing in Ecology and Conservation , 2022. [Paper ]
	Variational Feature Disentangling for Fine-Grained Few-Shot Classification Jingyi Xu, Hieu Le, Mingzhen Huang, ShahRukh Athar, Dimitris Samaras. IEEE International Conference on Computer Vision (ICCV), 2021. [PyTorch Code]
	Aerial-trained deep learning networks for surveying cetaceans from satellite imagery Alex Borowicz, Hieu Le, Grant Humphries, Georg Nehls, Caroline Hoschle, Vladislav Kosarev, Heather J. Lynch Plos One , 2019. [Paper ]
	From Shadow Segmentation to Shadow Removal Hieu Le, Dimitris Samaras. European Conference on Computer Vision (ECCV), 2020. [ Paper Project Page ]
	Shadow Removal via Shadow Image Decomposition Hieu Le, Dimitris Samaras. International Conference on Computer Vision (ICCV), 2019. [Paper Project Page ]
	Weakly Labeling the Antarctic: The Penguin Colony Case Hieu Le, Bento Gonçalves, Dimitris Samaras, Heather Lynch. IEEE Conference on Computer Vision and Pattern Recognition - Workshop (CVPRW), 2019. [Paper ]
	A+D Net: Training a Shadow Detector with Adversarial Shadow Attenuation. Hieu Le, Tomas F. Yago Vicente, Vu Nguyen, Minh Hoai, Dimitris Samaras. European Conference on Computer Vision (ECCV), 2018. [Paper]
	Iterative Crowd Counting. Viresh Ranjan, Hieu Le, and Minh Hoai. European Conference on Computer Vision (ECCV), 2018. [Paper]
	Co-localization with Category-Consistent CNN Features and Geodesic Distance Propagation Hieu Le, Chen-Ping Yu, Gregory Zelinsky, Dimitris Samaras. International Conference on Computer Vision - Workshop (ICCVW), 2017. [Paper]
	Geodesic Distance Histogram Feature for Video Segmentation Hieu Le, Vu Nguyen, Chen-Ping Yu, Dimitris Samaras Asian Conference on Computer Vision (ACCV), 2016. [link] [Paper] [Poster]
	Efficient video segmentation using parametric graph partitioning Chen-Ping Yu, Hieu Le, Gregory Zelinsky, Dimitris Samaras International Conference on Computer Vision (ICCV), 2015. [ paper]

	Learning Frame-Wise Emotion Intensity for Audio-Driven Talking-Head Generation Jingyi Xu, Hieu Le, Zhixin Shu, Yang Wang, Yi-Hsuan Tsai, Dimitris Samaras, 2025 [Preprint] We introduce a novel method for generating talking-head videos driven by audio input, incorporating frame-wise emotion intensity to enhance realism and expressiveness in visual output. The intensity is pseudo-labeled and we introduce a latent space that facilitates generating videos with varied frame-wise intensities.
	Instance-Aware Generalized Referring Expression Segmentation E-Ro Nguyen, Hieu Le, Dimitris Samaras, Michael Ryoo, 2025 [Preprint] [Project Page] We introduces a novel approach to referring expression segmentation with instance-awareness, automatically finding and linking object instances in the image with the textual entities describing them.
	Importance-based Token Merging for Diffusion Models Haoyu Wu, Jingyi Xu, Hieu Le, Dimitris Samaras, 2025 [Preprint] [Project Page] TL;DR: A novel token-merging method speeds up diffusion models by preserving important tokens, enhancing sample quality with minimal computational cost.
	QT-DoG: Quantization-Aware Training for Domain Generalization Saqib Javed, Hieu Le, Mathieu Salzmann, 2025 [Preprint] [Project Page] [Code] TL;DR: QT-DoG uses weight quantization as a regularizer to encourage flatter minima, enhancing domain generalization. It combines high accuracy with reduced model size and works well across datasets and architectures.
	Counting Stacked Objects from Multi-View Images Corentin Dumery, Noa Etté, Jingyi Xu, Aoxiang Fan, Ren Li, Hieu Le, Pascal Fua, 2025 [Preprint] TL;DR: We do 3D volume reconstruction from multi-view images to count hidden objects
	MedTet: An Online Motion Model for 4D Heart Reconstruction Yihong Chen, Jiancheng Yang, Deniz Sayin Mercadier, Hieu Le, Pascal Fua, 2025 [Preprint] [Code] [Project Page] TL;DR: This paper introduces a versatile framework for reconstructing 3D cardiac motion from limited real-time data, such as 2D slices or even 1D signals. Using a deformable tetrahedral grid, the method ensures anatomically consistent reconstructions suitable for intraoperative scenarios.