The aim of the AUS Lab I and Aus Lab II is to develop novel methods for complex problems for autonomous driving.  The required working hours are 120. The projects are supervised by professors and/or PhD candidates. 

At the end, a project report should be written, and a final workshop will be organized in which each participant will present their work in 10 minutes. The grade is given in the last week of the exam’s period. 

For complex problems, the laboratory work can be continued as an MSc thesis.

A very important remark is that generative artificial intelligence (ChatGPT, Gemini, etc.) can be used but you must mark in the report how you did use it.

We offer the students several topics, but own ideas can also be implemented.

Topics in the fall semester of 2026

Lidar-base perception supervised by Tekla Tóth:

https://cv.inf.elte.hu/wp-content/uploads/2026/02/Lab-Assignments-2025-2026-02.pdf

Underwater Motion Analysis and Reconstruction of Competitive Swimmers 
We are looking for a motivated student to join a research project focused on the computer-vision-based analysis of underwater videos of competitive swimmers. The aim of the project is to extract precise, quantitative motion information from real-world underwater recordings captured in challenging visual conditions. 
The project focuses on measuring and reconstructing key biomechanical and kinematic properties of swimmers, including body posture (positions and joint angles), swimming dynamics such as velocities and stroke frequencies, trajectories of individual body parts, and both instantaneous and time-averaged motion descriptors. These measurements enable a detailed, objective analysis of swimming technique and performance. 
From a technical perspective, the work combines computer vision, image and signal processing, and applied geometry. Core tasks include robust detection and segmentation of the swimmer’s body under water, handling visual artifacts such as bubbles and surface reflections, tracking body parts over time, estimating motion trajectories, and extracting physically meaningful signals from complex measurements. 
The project is strongly hands-on and data-driven. The student will work with real underwater video data, develop and test algorithms in Python, and create visualizations that support both scientific analysis and feedback for athletes and coaches. The work addresses open research problems and offers the opportunity to contribute to publications or thesis work. 

Image Rectification for Stereo Vision

Image rectification is a transformation process used to simplify the correspondence problem in stereo vision. When two cameras are mounted on a platform like the ELTECart, they are rarely perfectly aligned; slight rotations or vertical offsets mean that a point in the 3D world will appear at different vertical coordinates in each image. Rectification mathematically projects both images onto a common image plane such that all corresponding “epipolar lines” become horizontal and perfectly aligned. This ensures that for any pixel in the left image, the matching pixel in the right image is guaranteed to lie on the exact same horizontal row (scanline), effectively reducing a 2D search space to a 1D search. 

In a practical laboratory setting, this process relies on accurate extrinsic and intrinsic camera calibration. Students must first determine the relative rotation R and translation t between the two ELTECart cameras using a calibration pattern. Using these parameters, a rectification transform is computed to warp the raw, distorted images into a corrected pair. 

A fully automatic software should be developed by the student that estimates the rectifying transformations based on chessboard images. Then another software should be developed that rectifies the recorded images sequences for the stereo setup of ELTECar. 

Dense Stereo Matching and Disparity Estimation 

Once the ELTECart’s stereo images are rectified, the next challenge is dense matching, the process of finding a corresponding pixel in the right image for every single pixel in the left image. Unlike “sparse” matching, which only tracks distinct features like corners or edges, dense matching aims to produce a full disparity map. Students will implement or utilize algorithms—ranging from local block-matching techniques to global optimization methods like Semi-Global Matching (SGM)—to compute the horizontal shift (disparity) between corresponding points. This requires balancing matching accuracy against computational cost, as the algorithm must handle difficult scenarios such as textureless surfaces (like a smooth floor), repetitive patterns, and occlusions where a point is visible to only one camera. 

The output of this task is a dense disparity map, where the intensity of each pixel represents the distance shift between the two views. Because the cameras are rectified, this disparity is inversely proportional to the actual depth of objects in the scene. By leveraging the known baseline and focal length of the ELTECart’s sensors, students will transform this disparity map into a 3D point cloud. This  allows the vehicle to perceive the world in three dimensions, enabling downstream tasks like obstacle detection, path planning, and environment reconstruction. 

Machine learning is also very efficient for dense reconstruction, the student should compare the implemented methods to state-of-the-art pre-trained neural networks. 

Point-cloud rendering for visualisation

ELTECar is the vehicle of the Faculty of Informatics, mounted with several different sensors: 

• Digital cameras with normal and fisheye lenses 
• 3D LiDAR
• GPS device with RTK correction, reaching 3cm precision for localization, and 0.2° heading

The goal of this project is to make an application that turns 3D pointclouds into rendered pointclouds, and saves them into picture files. Those files then will be used together with the cameras and other data for data visualization.
The input data is .pcd formatted pointclouds (which will contain one single rotation of our LiDAR), the output is to be some kind of picture file (preferably .jpg) that has the pointcloud visualized, from a configurable perspective (not just BEV).
Optionally, it can be expanded to include datastreams from other sensors, for example GPS data visualized on a map.

Example visualization. Top: four cameras stitched to each other. Bottom: LiDAR points drawn to OpenStreetMap, ground points visualized by red. 

3D Vision via Affine transformations

The application of affine correspondences is beneficial over simple point correspondences in 3D vision because they capture more information about the local geometric transformation between images. While a simple point correspondence only matches the location (x,y) of a feature, an affine correspondence also matches the shape of the local neighborhood around that feature, effectively describing how a small circular region in one image is transformed into an ellipse in the other. This captures local rotation, scaling, and shear. This richer information provides more powerful constraints for solving 3D vision problems. It leads to more robust and accurate estimations, especially when dealing with significant changes in viewpoint where the apparent shape of objects is distorted.

Affine correspondences can be exploited for several 3D vision problems, like

• Image stitching/homography estimation
• Fudamental/essential matrix estimation
• Relative pose estimation
• Surface normal estimation.

All the methods are very sensitive to the quality of the affine transformations, retrieved from image pairs. The aim of the project is to compare different techniques and method both classical image processing ones and neural networks, in order to find the most accurate.

If the student is open to do research work, development of new solvers is also an option. In that case, after successful research, publication on international conferences/journals together with the supervisor is welcome and encouraged.

Comparison of Structure from Motion, Gaussian Splatting, and VGGT

Recent advancements in computer vision have introduced powerful techniques for 3D scene reconstruction, each with distinct strengths in processing speed and visual fidelity. This project focuses on evaluating and comparing three prominent methodologies:
traditional Structure from Motion (SfM), the novel 3D Gaussian Splatting, and VGGT. The primary objective is to systematically analyze their performance, rendering quality, and computational efficiency in generating accurate 3D models directly from standard 2D video sequences.
To ensure a comprehensive evaluation, the comparative analysis will be conducted using two distinct real-world recording scenarios. The first dataset will consist of indoor environments captured via handheld mobile video recordings, challenging the algorithms with complex lighting, close-range geometry, and irregular camera movements. The second scenario will utilize outdoor footage captured by vehicle-mounted cameras to assess the technologies’ scalability, robustness, and trajectory estimation in dynamic, large-scale environments. By processing these diverse datasets through all three pipelines, the research will benchmark their practical capabilities and overall suitability for different 3D reconstruction applications.

GPU exploitation for real-time 3D reconstruction and map

Achieving real-time 3D reconstruction and mapping remains a significant computational challenge, largely due to the massive volume of visual data and the complex optimization algorithms required. This project aims to investigate, implement, and evaluate advanced
GPU exploitation strategies to dramatically accelerate these computer vision pipelines. By leveraging modern GPU architectures, parallel processing paradigms (such as CUDA), and optimized memory management, the research will focus on overcoming traditional
bottlenecks to reduce processing times without compromising spatial accuracy or visual fidelity.
The scope of the project is technology-agnostic, allowing for the exploration of various 3D reconstruction methodologies. This could range from identifying bottlenecks and extending established photogrammetry pipelines—such as accelerating feature
extraction, matching, or bundle adjustment in COLMAP—to optimizing newer neural, volumetric, or SLAM-based approaches. The primary task is to apply targeted hardware acceleration techniques and benchmark the resulting performance against standard implementations, ultimately demonstrating a clear pathway toward robust, real-time mapping solutions suitable for autonomous navigation, robotics, or augmented reality.

Affine Transformations for SfM/SLAM

Traditional Structure from Motion (SfM) and Simultaneous Localization and Mapping (SLAM) systems typically output sparse 3D point clouds that lack explicit information about the underlying surface geometry. This research project focuses on exploiting local
affine transformations between image pairs to extract significantly richer geometric data. By mathematically analyzing how local image patches deform between views—based on known or estimated affine transformations—it is possible to compute not only the precise
3D coordinates of a point but also its corresponding surface normal. This novel approach directly yields the orientation of the observed surfaces, seamlessly bridging the gap between sparse point extraction and dense surface reconstruction.
As an integral part of an ongoing, specialized research initiative, this project offers students the opportunity to dive deep into projective geometry and advanced computer vision algorithms.

The primary objective is to implement, refine, and test these affine-based mathematical models to reliably estimate normal vectors within standard SfM/SLAM frameworks. Students will be tasked with validating the methodology using diverse image pairs, analyzing the accuracy of the computed surface normals, and exploring how this enriched spatial data can improve subsequent tasks such as dense meshing or robust camera tracking. This is a unique opportunity to contribute directly to cutting-edge research and expand the theoretical boundaries of 3D vision.

Surface Reconstruction from 3D models after Gaussian Splatting

While 3D Gaussian Splatting (3DGS) has recently revolutionized novel view synthesis with
its exceptional rendering speed and visual quality, its underlying scene representation
relies on unstructured volumetric ellipsoids rather than explicit geometric surfaces. This
project addresses the critical challenge of deriving accurate, continuous 3D surface
meshes from trained 3DGS models. The primary objective is to investigate and implement
state-of-the-art surface extraction techniques, bridging the gap between high-fidelity
volumetric radiance fields and traditional polygon-based 3D geometry, which remains
essential for physical simulations, collision detection, and standard rendering engines.
To rigorously evaluate the effectiveness of the surface reconstruction methods, the
research will utilize video data from two contrasting real-world scenarios. The first dataset
will feature complex indoor environments captured via handheld mobile devices, testing
the algorithms against variable lighting, occlusions, and detailed close-range geometries.
The second scenario will rely on large-scale outdoor footage recorded by vehicle-mounted
cameras to assess the robustness, scalability, and mesh completeness along continuous,
dynamic trajectories. By applying the surface extraction pipelines to these diverse
environments, the project will benchmark the geometric precision and overall quality of
meshes generated directly from Gaussian splats.

Motion Blur Removal Using SfM Geometry

Motion blur is an inherent limitation of video captured under motion, and it is
conventionally treated as a blind image restoration problem in which both the sharp image
and the unknown blur kernel must be recovered simultaneously. This project takes a
different route by exploiting the fact that a Structure from Motion pipeline already recovers
precisely the quantities that generate the blur: the camera trajectory, the intrinsic
parameters, and after dense reconstruction the scene depth. Since a blurred frame can be
modelled as the integral of the sharp image over the exposure interval, warped by the timevarying camera pose, these estimates allow the spatially varying blur kernel to be
synthesised directly rather than guessed. The primary objective is to implement a pipeline
that derives per-pixel blur kernels from SfM output and applies non-blind deconvolution,
thereby converting a severely ill-posed blind problem into a geometrically constrained one.
The project begins with a thorough exploration of the topic and a survey of the current state
of the art, covering both classical geometry-driven deblurring and recent approaches that
jointly optimise scene representation and camera motion. Students are then expected to
select representative methods, reproduce their results on real recordings, and evaluate
them critically, identifying the assumptions on which they depend and the conditions
under which they break down