Subject: Deep Neural Networks for Point-Cloud Scene Analysis
Lab: LIP6, CNRS Sorbonne University
Point cloud analysis constitutes a major challenge in computer vision, leveraging high-volume 3D geometric data from LiDAR and RGB-D sensors for comprehensive spatial and semantic scene understanding. This capability is mandatory across demanding, real-world applications such as autonomous perception and large-scale 3D environment reconstruction, necessitating robust, low-latency representation learning. While deep neural networks achieve state-of-the-art performance for tasks like segmentation and detection, this success is predicated on large-scale, annotated datasets, establishing a critical data scarcity bottleneck. This challenge is amplified in 4D spatio-temporal sequences, where enforcing temporal consistency in annotations incurs prohibitive labeling overhead. To mitigate these issues, 3D Foundation Models (FMs) facilitate knowledge transfer, yet the inherent scale and the non-Euclidean nature of point clouds necessitate complex, permutation-invariant architectures. This results in high computational complexity and substantial memory footprints. This architectural overhead directly impedes deployment on resource-constrained edge devices, often leading to untenable inference latency and degradation of geometric fidelity. Therefore, managing labeled data scarcity and mitigating severe model overhead remain the primary hurdles to widespread, robust 3D vision inference.
To address the aforementioned critical bottlenecks of limited labeled data and excessive model overhead, this research subject proposes an integrated methodology combining novel self-supervised learning (SSL) and active learning (AL) paradigms with architectural optimization and compression techniques, ensuring the resulting 3D FMs achieve robustness, accuracy, and efficient inference. These objectives encompass:
- Self-Supervised Learning (SSL): The goal is to formalize novel SSL frameworks and pretext tasks designed to leverage the inherent spatio-temporal structural regularities present in large-scale unlabeled 3D and 4D point-cloud data.
- Active Learning & PEFT: The objective is to develop principled active learning strategies leveraging uncertainty and density metrics to facilitate frugal annotation. This also relies on Parameter-Efficient Fine-Tuning (PEFT) techniques (e.g., prompt tuning, LoRA) to streamline the adaptation of FMs.
- Resource-Aware Neural Architectures: The goal is to design optimized, resource-aware neural architectures and point-cloud representations. Training will utilize task-appropriate perceptual quality losses to enhance geometric and semantic fidelity across various machine vision and representation learning tasks (recognition, compression, etc).
Keywords: deep neural networks, generative models, point cloud analysis, machine vision
Starting and Duration: the position may start as early as March/April 2026 for a duration of six months. This could be followed by a PhD position.
Lab: LIP6, CNRS, Sorbonne University, Paris.
Background: we are seeking a highly motivated candidate, with a preferred background in applied mathematics or computer science with more emphasis on statistics, machine / deep learning and 2D/3D visual data processing, and familiar with existing machine learning tools and programming platforms.
Applications: should be sent to “hichem [dot] sahbi [at] sorbonne-universite [dot] fr” including a CV and all the available university studies transcripts (and if available recommendation letters).
