PoseShot: hybrid CNN–BiLSTM transformer model for free throw action recognition via pose analysis

Smarter Practice for a Classic Basketball Shot

For anyone who has ever stepped to the free throw line, the difference between a swish and a miss can feel mysterious. This study shows how artificial intelligence can turn that mystery into clear, objective feedback. By carefully reading a player’s body posture during a free throw, the researchers’ system, called PoseShot, breaks the shot into understandable stages and reveals which movements separate a solid technique from a shaky one.

Breaking a Free Throw into Clear Stages

Instead of treating a free throw as a single, simple action, PoseShot views it as a sequence of five distinct phases: dribble, hold, raise, throw, and follow-through. The team collected 75 videos of players performing free throws in a variety of gyms, lighting conditions, and camera angles, then labeled thousands of frames with these phase names. This fine-grained approach lets the system ask much more specific questions: Is the player pausing long enough in the hold phase? Does the arm path look smooth during the raise? Is the follow-through consistent from shot to shot?

Reading the Body with Computer Vision

To see what the player is doing, PoseShot combines two complementary views of the same motion. First, it looks at the raw video images and uses a type of neural network known as a convolutional network to capture shapes, limb positions, and motion cues across each frame. Second, it runs the video through a pose-estimation tool that marks 2D joint locations throughout the body. From these points it computes eight key joint angles, including both elbows, hips, and knees. These angles form a kind of numeric fingerprint of the player’s posture through time, capturing how energy travels from the legs, through the torso, to the shooting arm.

Blending Two Streams of Motion Intelligence

The heart of PoseShot is a dual-path architecture that processes pictures and posture side by side. The image stream flows through several layers that gradually condense visual details into compact features. The posture stream feeds the sequence of joint angles into a special time-aware network that tracks how one body position leads to the next, both forward and backward in time. On top of both streams, the system adds transformer layers that act like attention mechanisms, highlighting the most informative moments and relationships in the motion. Finally, it merges the two streams and sends them into a classifier that decides, frame by frame, which phase of the free throw the player is in.

Outperforming Other AI Coaches

To judge how well PoseShot works, the researchers compared it with a range of popular deep-learning models, from classic image networks to modern transformer-based designs. Using a balanced metric that rewards both accurate hits and avoiding mistakes, PoseShot scored about 96 out of 100, beating all competing methods overall and especially excelling at the crucial throw phase. The system was particularly strong at recognizing clear-cut movements like hold, throw, and follow-through. The hardest distinctions were between dribble and raise, which often blend into one another in real play and differ in subtle timing and speed rather than obvious posture changes.

From Data to Actionable Coaching

By turning high-speed video and joint measurements into reliable phase labels, PoseShot offers coaches more than just a highlight reel. It provides a structured way to see where a player’s form breaks down, how consistently they repeat key positions, and which parts of the motion separate made shots from misses. Although the system still struggles with very similar movements and currently runs best on powerful computers, the study points toward a future in which players at all levels can receive instant, objective breakdowns of their technique. In practical terms, that means more targeted practice, fewer bad habits, and a clearer path from the free throw line to the bottom of the net.

Citation: Hsu, WC., Lee, CC., Lee, YH. et al. PoseShot: hybrid CNN–BiLSTM transformer model for free throw action recognition via pose analysis. Sci Rep 16, 11478 (2026). https://doi.org/10.1038/s41598-026-41025-0

Keywords: basketball free throw, sports analytics, motion capture, deep learning, action recognition