TrackNet: A Deep Learning Network for Tracking High-speed and Tiny Objects in Sports Applications
Paper • 1907.03698 • Published • 2
TrackNet V1 — Tennis Ball Tracking
TrackNet V1 weights for high-speed tennis ball detection from broadcast video. The model takes 3 consecutive frames (resized to 640x360) and outputs a probability heatmap of the ball position.
Architecture
- Input: 3 consecutive RGB frames (640x360) = 9-channel input
- Output: 640x360 heatmap (ball probability)
- Backbone: Custom CNN with encoder-decoder architecture (DeconvNet-style)
Usage
import torch
from tracknet import BallDetectorV2
detector = BallDetectorV2(path_model="tracknet_weights.pth", device="mps")
ball_positions = detector.infer_model(frames, use_frame_diff=True)
Credits
- TrackNet by Yu-Chuan Huang, I-No Liao, Ching-Hsuan Chen, Tsui-Kang Ik, Wen-Chih Peng. Paper: TrackNet: A Deep Learning Network for Tracking High-speed and Tiny Objects in Sports Applications (AVSS 2019).
- Implementation adapted from yastrebksv/TrackNet and yastrebksv/TennisProject. Enhanced with frame-differencing and Kalman smoothing for this project.
Part of
This model is part of the Sports Commentator from Video project — an automated tennis commentary pipeline using TrackNet, CatBoost, VideoMAE, Qwen2.5, RF-DETR, and F5-TTS.