Joakim Bruslund Haurum

I am a Postdoctoral Researcher in the Visual Analysis and Perception Laboratory at Aalborg University, where I work on computer vision and fine-grained visual categorization. I'm also an affiliated member of the Pioneer Centre for Artificial Intelligence and an ELLIS Member. Previously, I was a member of the Young Academy Panel at the Danish Data Science Academy.

During my Ph.D. I worked on Computer Vision Aided Sewer Inspections and Marine Vision, advised by Professor Thomas B. Moeslund. I have previously visited the Human Pose Recovery and Behavior Analysis in 2021 working with Professor Sergio Escalera, been a visiting researcher at the Vector Institute and the Machine Learning Research Group at the University of Guelph in 2023 working with Professor Graham Taylor, and in 2024 I was a visiting researcher at the University of Edinburgh working together with Reader Oisin Mac Aodha.

Email  /  Google Scholar  /  Github  /  HuggingFace  /  Twitter  /  LinkedIn  /  Headshot

I'm interested in computer vision, machine learning, marine vision, fine-grained visual categorization, and real-life applications. Representative papers are highlighted.

We explore whether pretrained models can provide a useful representation space for datasets they were not trained on for the purpose of grouping novel unlabelled data into meaningful clusters.

We introduce CLIBD, the first approach to use contrastive learning for aligning biological images with DNA barcodes and taxonomic labels to enhance taxonomic classification.

We construct a large-scale fine-grained dataset of the Insect class with 5M data samples, each containing a biological taxonomic annotation, DNA barcode sequence, geographical location, and RGB image.

We present Agglomerative Token Clustering (ATC), a novel token merging method that consistently outperforms previous token merging and pruning methods across image classification, image synthesis, and object detection & segmentation tasks.

We investigate whether using 3D Gaussian Splatting (3DGS) instead of NeRFs to produce stereo camera views leads to better dense disparity labels.

We investigate the effect of integrating a deblurring module with a saliency predictor to perform better point sampling for direct visual odometry.

We propose AssemblyNet, a novel dataset for predicting part directions in assembly models for exploded view visuzalitions. We propose a novel two-path network for predicting part directions.

We propose BarcodeBERT, the first self-supervised method for general biodiversity analysis and highlight how the role of self-supervised pretraining in achieving high-accuracy DNA barcode-based identification at the species and genus level.

We construct a large-scale fine-grained dataset of the Insect class with 1M data samples, each containing a biological taxonomic annotation, DNA barcode sequence, and RGB image.

We conduct the first systematic comparison and analysis of 10 state-of-the-art token reduction methods across four image classification datasets, trained using a single codebase and consistent training protocol.

We propose a novel method for to model spatial semantics in images, where features are aggregated at different scales non-locally using a lightweight vision transformer, and novel Sinkhorn clustering-based tokenizer.

We propose a set of metrics to determine the complexity of Multi-Object Tracking datasets along three axes: Motion, Appearance, and Occlusion.

My Ph.D. thesis summarizing my work on computer vision aided sewer inspections. Supervised by Thomas B. Moeslund.

Assessment committe: Georgios A. Triantafyllidi (chair.), Serge Belongie, and Graham W. Taylor.

We propose a keypoint matching based framework for re-identification of Giant Sunfish, comparing SIFT, RootSIFT, and SuperPoint descriptors.

We propose a novel decoder-focused multi-task classification architecture called Cross-Task Graph Neural Network (CT-GNN), which refines the disjointed per-task predictions using cross-task information.

We present the world's first open-source sewer defect dataset, with multi-label annotations and 1.3 million images from Danish sewers.

Investigation into comibning synthetic and real point cloud data of sewer pipes for sewer defect classification using PointNet and DGCNN.

We estimate the water level in sewer pipes by casting the task as a classification and regression problem.

We presenta a dataset for 3D multi-object tracking of zebrafish, captured using an off-the-shelf setup.

We propose a system for generating synthetic point clouds of sewer pipes using Structured Domain Randomization for the generation of the sewer systems and an approximated model of a Pico Flexx Time-of-Flight camera.

We reviewed 113 articles on image-based automated sewer inspection methodologies, finding that there are is a severe lack of 1) publicly available datasets, 2) commonly agreed upon evaluation protocols, and 3) open-sourced code.

We show that it is possible to re-identify zebrafish from a sideview, using metric learning and classical feature descriptors.

We present a new publicly available underwater dataset with annotated image sequences of fish, crabs, and starfish captured in brackish water with varying visibility.

We design a system for the detection of rainfall by the use of surveillance cameras, and compare it against the former state-of-the-art method for rain detection on our new AAU Visual Rain Dataset (VIRADA).

We apply pixel reprojection on nine 360 degree renderings to enable 3D motion and introduce motion parallax effects, without explicit knowledge of the 3D geometry.

We evaluate and discuss the experiences acquired during the 18 month long production, qualification and testing phase of the AAUSAT5 cubesat.

This website layout was adapted from Jon Barron's template.