Latest writing

Writing

11 August 2026 · 3 min read

Thermal imaging for detection work: what changes when light becomes heat

A thermal camera does not see light, it sees temperature. That breaks half of your RGB intuition. What I had to relearn to do detection on the thermal channel.

4 August 2026 · 3 min read

The gap between the notebook and the board

A detector that scores well on a workstation and fails on an embedded target is usually not a model problem. Four places the gap actually opens up.

28 July 2026 · 3 min read

Why I labelled it a shot candidate and not a goal

Building event detection from ball trajectory on uncalibrated footage, and why naming the limitation in the output was cheaper than pretending it was not there.

21 July 2026 · 3 min read

NMS and confidence thresholds: the last, underrated step of detection

Your detector outputs thousands of candidate boxes per frame. Two cheap post-processing choices decide what anyone actually sees, and they get far less attention than they deserve.

14 July 2026 · 4 min read

The dataset decisions that decide your mAP before training starts

Tiling overlap, leakage-free splits and class-balanced augmentation did more for our aerial detector than any hyperparameter. Here is what we changed and why.

30 June 2026 · 7 min read

Image stitching and orthomosaics: turning frames into a map

How a few hundred overlapping drone frames become one image, why a panorama is not a map, and the flight-planning decisions that matter more than the algorithm.

16 June 2026 · 3 min read

Camera calibration and lens distortion: why your straight lines bend

The runway is straight, the image says otherwise. What calibration actually estimates, how I run the checkerboard ritual, and where uncorrected distortion bites a mapping pipeline.

2 June 2026 · 5 min read

ROS 2 for vision systems: nodes, topics and what actually matters

Notes from putting ROS 2 on the table while designing a SUAS mission architecture: what nodes, topics and QoS buy a vision pipeline, and when a small system does not need them.

19 May 2026 · 7 min read

The Kalman filter without the matrix fog

Every tracker I have shipped sits on a Kalman filter, and none of them needed the full matrix derivation to work. Here is the intuition that actually matters.

5 May 2026 · 7 min read

Object tracking and ByteTrack: keeping identity across frames

A detector tells you what is in a frame. Tracking tells you it is the same thing as last frame. Why that matters, and why ByteTrack's one idea works so well.

7 April 2026 · 8 min read

TensorRT and FP16: getting a trained model onto real hardware

A trained checkpoint is not a product. What TensorRT actually does, why FP16 is usually free, and the checks I run before I trust an engine on a Jetson.

10 March 2026 · 3 min read

Detecting animal disease with computer vision: a field worth building in

Every hard problem in livestock health monitoring is one I already work on: tracking, thermal imaging, ugly data, edge deployment. Notes on why this field deserves more engineers.

10 February 2026 · 3 min read

Can you do computer vision on a VR headset?

A standalone headset already runs more computer vision than most robots. The real question is how much of that your own code gets to use, and what it costs.

13 January 2026 · 3 min read

Developing for Oculus/Meta Quest: notes from shipped projects

The Quest is a mobile chip strapped to your face. Frame rate is a contract, the editor lies, and comfort cannot be patched in later. Notes from shipped training projects.

9 December 2025 · 3 min read

VR, AR, XR: what the letters mean when you build for them

In meetings the three letters get used interchangeably. In the engine they decide who renders the world, what your frame budget is and what failure looks like.

4 November 2025 · 7 min read

OpenMP vs CUDA: two kinds of parallel, one real-time budget

I use OpenMP and CUDA in the same real-time vision pipeline. Which one fits which job, why data movement is the real bill, and who actually eats your frame budget.

14 October 2025 · 3 min read

CUDA, CUDA C and the OpenCV CUDA module: which one are you actually using?

Three different things share the CUDA name and people mix them up daily. The platform, the language and the library have different costs, and picking the wrong layer wastes weeks.

16 September 2025 · 3 min read

Setting up an NVIDIA Jetson: JetPack, first boot and the road to TensorRT

Every Jetson project I have worked on lost days to setup before it processed a single frame. Here is the order that stops the leak: version choice, power mode, the right wheels, then TensorRT.

12 August 2025 · 3 min read

C++ multithreading for real-time pipelines: the parts that bite

Capture, inference, tracking and a Qt interface, all sharing one machine. What I learned about threads from keeping a detection pipeline honest under a real-time budget.

8 July 2025 · 3 min read

What the FFT actually does, and what it does for images

The Fourier transform is not a formula to memorise, it is a change of viewpoint. What that viewpoint buys you in a real vision pipeline, and where it does not help at all.

10 June 2025 · 8 min read

The image processing fundamentals I still use every week

Thresholding, morphology and contours built my first robot, and they still carry my detection pipelines. A tour of the classics that refuse to retire.

13 May 2025 · 3 min read

Git and GitHub: the five habits that actually matter

Not forty commands. Five habits: commit messages that record why, one branch per experiment, data out of the repo, reading your own diff, and tagging what ships.

15 April 2025 · 3 min read

Building operator interfaces in Qt: lessons from the field

The model gets the attention, but the operator screen decides whether the system is usable. What I learned building Qt interfaces for real-time tracking systems.