Register now After registration you will be able to apply for this opportunity online.
This opportunity is not published. No applications will be accepted.
Active Sensing with Diffusion-based Motion Generation
The efficacy of the diffusion model has been demonstrated across various computer vision applications, notably in image generation and editing[1][2]. This thesis aims to extend its generative capabilities to the domain of active sensing, specifically facilitating a mobile robot's autonomous exploration and mapping of its environment. Current methods for active sensing and viewpoint selection predominantly lean on either volumetric reconstruction, which necessitates manually crafted metrics and is bound by the reconstruction method's limitations, or reinforcement learning, which demands significant training efforts and often struggles with generalization. We anticipate that adopting a diffusion-based approach will surpass these constraints and lead to enhancements in the field.
Keywords: Diffusion model, active sensing, deep learning
The efficacy of the diffusion model has been demonstrated across various computer vision applications, notably in image generation and editing[1][2]. This thesis aims to extend its generative capabilities to the domain of active sensing, specifically facilitating a mobile robot's autonomous exploration and mapping of its environment. Current methods for active sensing and viewpoint selection predominantly lean on either volumetric reconstruction, which necessitates manually crafted metrics and is bound by the reconstruction method's limitations, or reinforcement learning, which demands significant training efforts and often struggles with generalization. We anticipate that adopting a diffusion-based approach will surpass these constraints and lead to enhancements in the field.
[1] Yu et al., arxiv 2024, Motion-I2V: Consistent and Controllable Image-to-Video Generation with Explicit Motion Modeling,
[2] Geng et al., ICLR 2024, Motion Guidance: Diffusion-Based Image Editing with Differentiable Motion Estimators
The efficacy of the diffusion model has been demonstrated across various computer vision applications, notably in image generation and editing[1][2]. This thesis aims to extend its generative capabilities to the domain of active sensing, specifically facilitating a mobile robot's autonomous exploration and mapping of its environment. Current methods for active sensing and viewpoint selection predominantly lean on either volumetric reconstruction, which necessitates manually crafted metrics and is bound by the reconstruction method's limitations, or reinforcement learning, which demands significant training efforts and often struggles with generalization. We anticipate that adopting a diffusion-based approach will surpass these constraints and lead to enhancements in the field.
[1] Yu et al., arxiv 2024, Motion-I2V: Consistent and Controllable Image-to-Video Generation with Explicit Motion Modeling, [2] Geng et al., ICLR 2024, Motion Guidance: Diffusion-Based Image Editing with Differentiable Motion Estimators
Build an active sensing/viewpoint selection pipeline using diffusion-based motion generation pipeline
Build an active sensing/viewpoint selection pipeline using diffusion-based motion generation pipeline