Ruku

Swapped my ball tracker from TrackNet (3-frame temporal CNN) to RF-DETR (transformer, single-frame). Trained on ~4,700 labeled pickleball frames from public data. Detection rate jumped from 47% to 86%.

@mirrash7 @roboflow Great read!

My first blog post for @roboflow is live 🎉 The new 'Track Class Lock' block reduces class flickering and turns per-frame guesses into settled object identities blog.roboflow.com/reduce-class-f…

@skalskip92 Looks awesome. Will def compare it to the torchvision Keypoint R-CNN I ran for my pickleball players skeletons

RF-DETR keypoints is finally out preview release: real-time transformer keypoint detection Apache 2.0 71.8 AP on COCO, 9.7ms on T4. outperforms YOLO11-pose and YOLO26-pose at similar latency

@josephofiowa Just ran it. SAM3 labeled 30 frames in minutes from a "boat" prompt. I added a "car" class and excluded it to keep the parking lots clean. Way less work than the 7 I did by hand, and the labels look great. Super impressive!

@lohba_ pretty mesmerizing. try SAM 3 autolabel --> dataset --> RF-DETR? docs.roboflow.com/annotate/ai-la…

Fine-tuned RF-DETR on just 7 labeled frames to detect boats in a packed marina from drone footage. 7 felt like way too few to be a real test but it worked surprisingly well.

Before vs after fine-tuning RF-DETR on just 7 labeled frames

@RyanDiew awesome, thanks for sharing!

@lohba_ grounding dino auto-labels the frames, then i distill it into a custom yolov11 that runs fast on-device. hard-example mining catches what it misses.

taught it to track every player. then gave it eyes. pose. ball. rim. one shaky clip from the bleachers. no tripod, no tags. the whole floor is data now 🫡

@skalskip92 Can’t wait 👀

looks like we are working on the same thing haha; more soon

infinite velodrome

8 seconds is enough time for a Polymarket degen to get a real edge in live sports betting

@Aman09143227 RF-DETR

@lohba_ For ball tracking what you use ?

Shots from one of my rec. pickleball games visualized in 3D: - Contact timing from audio - Player positions from vision - Ball flight reconstructed between contacts - All from a single iphone video

@samptampubolon appreciate it!

@lohba_ looks great man! amazing stuff.

3D body overlay on players with off-the-shelf HMR2.0

Adding ball-speed metrics. The ball is tracked, its flight fit to a ballistic arc, and each shot's pace clocked at the bounce. Shown as a blue-to-red meter for relative pace rather than a hard mph.

I ran 2D pose estimation on a pickleball rally, with skeletons on all four players through the entire point: - Keypoint R-CNN (torchvision) detects the 17 body joints for each player - PyTorch runs the model inference - OpenCV connects the joints and draws the skeleton overlay onto the frames

@StephanSturges thanks!

@lohba_ cool!

For each frame, I detect every player and map their feet from pixel space onto real court coordinates using the court homography, then tag each position with the correct player identity. You can accumulate all of a given player's positions into a two-dimensional grid laid over the court and color each cell by how often that player stood there, so that denser areas represent more time spent and the result is a per-player occupancy heatmap.

@kirbytheodor Neat, will be following your journey 👍🏻

@lohba_ a home server, mainly for local inference

best investment of my life