Before writing a single line of code in my bachelor’s project, I read about 35 pieces of work by the four cohorts before me. That’s the honest status after five weeks.

GorillaScout is my bachelor’s project at HPI, at Prof. Gerard de Melo’s chair for Artificial Intelligence and Intelligent Systems. We’re eight students, the project runs from the winter term until summer 2027, full-time in the summer. It’s about western lowland gorillas, which are critically endangered. The videos the previous cohorts work with come from camera traps in a national park in the Republic of the Congo. And camera traps produce more video than people can go through by hand.

HPI bachelor’s projects have worked on this for years, a new cohort every year. We’re the fifth. The cohort before last, GorillaWatch, even published a paper at WACV 2026. So our job isn’t to start from zero but to keep developing an existing pipeline. What exactly our part will be, and which role I take on in it, hasn’t been decided yet.

The pipeline can be explained in five steps. (The demo on this page is something else: a rebuild of the project’s web platform, which earlier cohorts built, with invented animals and drawn frames instead of real camera-trap footage.) One, detect: a model finds the gorillas in each frame and draws a box around them. Two, track: across several frames it becomes clear which box belongs to which animal, so single frames turn into tracks. Three, embed: each track is turned into a vector of numbers that describes what the animal looks like. Four, re-identify: similar vectors probably belong to the same individual, even in a different video. The hard part is that new animals can show up that nobody has ever seen. Five, count: how many different gorillas are there?

How did I get into it? Through those 35 earlier works: bachelor’s theses, project documentation and the WACV paper. Together with Claude I turned them into a searchable knowledge base and a study guide with 13 chapters, from detection through tracking to re-identification and the metrics used to evaluate it. Every statement has a source, and where two works contradict each other, that goes into a separate register. The last chapter is called “where you can start” and collects 14 places where the work could continue. The texts were written with AI assistance, and I spot-checked the numbers.

On top of that came getting into the chair’s infrastructure: GPU servers, SSH access, Docker, a Slurm cluster and experiment tracking. For the introduction session I wrote myself a study sheet.

What I’m taking away so far: you don’t enter a project with a history through code, you enter it by reading. If you don’t know the decisions of earlier cohorts, you’ll rebuild them by accident. And AI helps a lot with structuring 35 documents, but only if every statement points to a source.

Tools and skills so far: literature work, computer vision basics for detection, tracking and re-identification, evaluation metrics for open-set re-identification, Claude for structuring the material, GPU servers, SSH, Docker and Slurm.

The building starts now. This post will get a follow-up.