Evaldas Vaičekauskas
Selected Projects & Research
Vilnius, Lithuania ·Email ·LinkedIn ·ResearchGate
I design systems that extract structure from complex data and turn it into something people can use or experience. The input might be a 3D laser scan, or a painting — the arc is the same: find the features that carry meaning, model them so the result can be evaluated and understood, and turn that into something concrete: a robot trajectory, a saliency map, a piece of music. I am drawn especially to interpretable methods, where how a model reaches its output matters as much as the output itself.
Beyond the Black Box: An Interpretable Saliency Framework for Abstract Art
I chose this topic to answer questions I had been circling for a long time: where do people look when they observe an abstract painting, which elements draw the most attention, and which visual features interact most strongly with established aesthetic principles. A model that predicts human attention in abstract art becomes a way to study how a work is experienced over time, and how a hierarchy of significance emerges within it.
I engineered 35 visual heuristics derived from art and aesthetics theory, organized around long-standing principles such as contrast, isolation, symmetry, and grouping. A genetic algorithm optimized the weights of this heuristic set to maximize agreement between predicted and ground-truth human attention; because the optimization drove weak heuristics toward zero, it doubled as feature selection. After tuning against real attention data, each theory-defined group retained comparable overall significance — a model structured by century-old artistic principles, then calibrated numerically against how people actually look.
The framework's predictive accuracy was comparable to generic CNN-based saliency models, with one decisive difference: its composition is interpretable (explicit, inspectable feature weights) and its working principles are explainable (concrete, named heuristics rather than opaque activations). It matches the black box while remaining a glass box — which, in any setting where a prediction has to be trusted and justified, is the property that matters most.
The topic is far from exhausted: the correspondence between predicted and actual attention can still be tightened, and an emotional-response dimension would broaden the model's range — a direction I intend to keep developing.
3D-Driven Trajectory Generation for Robotic Machining
The challenge: automate the sanding of wooden furniture parts whose real geometry never matches the CAD model — so the system had to perceive each individual workpiece and generate a bespoke trajectory from the scan rather than run a fixed program.
I led the project and owned the trajectory-generation and robot-execution pipeline. In Python on the RoboDK API, I parameterised the sander contact poses into a transform chain and built a discretised feasibility map over it, where each cell encoded whether a valid robot configuration (an inverse-kinematics solution) existed. I then applied a safety-margin inflation to the infeasible regions and ran a custom multi-objective A* search — weighting path smoothness, safety clearance, and forward progression — to produce continuous, collision-aware trajectories, which I converted into joint paths and executable robot programs validated in RoboDK simulation.
These trajectories were driven by 3D perception. The contact poses came from a point-cloud pipeline (Open3D, NumPy) I co-developed with a colleague: it extracted part geometry from 3D scans by detecting right-angle edges from discontinuities in local surface-normal orientation, placing sander contact points at the midpoints between paired edges around the perimeter, and assigning each point a contact orientation from its local surface normal.
As R&D lead I authored the research grant proposal that funded the work, defined and scheduled the staged work packages, and owned deliverable reporting and deadlines across the engineering team, coordinating the perception and trajectory subsystems with a university research partner (VilniusTech). The system was designed to correct in real time against force-sensor feedback to hold constant contact pressure; my own contribution centred on the vision-driven trajectory generation and planning. I also completed COGNEX industrial vision certification.
Sonification of Visual Art
This grows directly out of the saliency work: I couldn't leave the model at prediction — I wanted to hear what it found. The pipeline reads a painting through the saliency model, extracting its attention centres with deliberately varied patch sizes to capture both sharp central fixation and softer peripheral vision. Each patch is run through a visual feature extractor — colour, texture, intensity, blob structure — and those features are mapped onto musical ones, which drive synthesizers (built on Supriya, the Python interface to SuperCollider) controlling pitch, brightness, modulation, and distortion. The synthesised voices are then sequenced into a composed piece whose structure, tempo, and foreground/background follow the structure of the painting itself.
The hard, open part is the visual-to-musical mapping, where there is no definitive answer and real artistic choice — I've been grounding my intuitions in the literature on how painters hear musicality in their work and how composers have been moved by paintings. The current pieces are experimental, but they make the mapping audible: you can hear the structure the model found in the image — watch a short demo on Instagram. I'm in early talks with an abstract painter in Vilnius about sonifying his works in the space where they're exhibited.
Underneath it is a question I keep returning to: whether there is something universal in how we experience art. My working hypothesis is that if that experience is fundamentally emotional, then for a given painting there exists a piece of music that evokes a corresponding emotional response — that emotion is what survives translation from one medium to another. The project is my way of testing that by building the bridge and listening to what crosses.