Posts on TechnoCurl

Tracking Curling Rocks, Part 4: Bootstrapping a Detection Model

Mon, 11 May 2026 08:00:00 -0500

Having laid the ground work in camera calibration (part 3) and with our initial exploration done using artificial neural networks models to do rock detection (part 2), we are now in a good position to train our custom models. The hope is that this will deliver better accuracy while improving efficiency i.e. faster detection.

We need to label some images to train models. But since I don’t have a team of professional data labellers at my disposal, I would need to make this process as easy as possible.

Tracking Curling Rocks, Part 3: New Camera System

Fri, 17 Apr 2026 08:00:00 -0500

While I was conducting my experiments in parts 1 and 2, the club planned for and got a new streaming system installed by Curling Stadium in the summer of 2025. Their standard setup puts 4 HD 1080p cameras on each sheet: two PTZ (pan-tilt-zoom) cameras mounted on the end walls for broadcast-style coverage, and two POV (point-of-view) overhead cameras mounted directly above each house. For my purposes, the overhead cameras are the interesting part – they give a fisheye view of the house end from directly above, which is exactly what I need for rock tracking. With six overhead cameras across three sheets, it was time to recalibrate everything.

Tracking Curling Rocks, Part 2: Off-the-Shelf Deep Learning

Sat, 04 Apr 2026 06:00:00 -0500

In part 1, I built a rock detection pipeline using classical computer vision – fisheye correction, HSV color matching, and the Hough circle transform. It worked reasonably well, but was held together by hand-tuned thresholds that would fall apart the moment the lighting changed or someone switched to a different camera. The hog line kept getting detected as a red rock. I wanted to see if off-the-shelf deep learning models could do better, without having to train anything myself.

Tracking Curling Rocks, Part 1: Classical Computer Vision

Wed, 25 Mar 2026 00:00:00 -0500

When I first started curling a few years ago, as a front end player, I found it difficult to keep track of the shots played and maintain a good mental picture of the end progression. The club does have monitors behind the hack to help the front end get a better view of the house, but between focusing on my then barely existent delivery consistency and keeping up with the pace of the game, I tended to lose track. I wondered at the time – if we could overlay the trajectory of the delivered stone and the other stones that were set in motion, that would be pretty helpful.

If you’ve watched curling on TV, you’ve seen something in this direction. Commentators use telestrators to draw lines and place stones on screen before a shot to discuss strategy, but that’s manual and only covers the planned shot. On the automated side, CurlingHunter (used at the 2022 Beijing Olympics) and SPECTO Curling (LIDAR-based) can actually track stones in real time as they move. I thought it would be fun to try building something like that myself.