Keyframe-Based Sports Video Segmentation with a Task-Specific PSPNet Adaptation: An Offline Evaluation for Decision Support

Sports-video segmentation is difficult because of rapid motion, occlusion, scale variation, and visually complex venues. This study evaluates a two-stage pipeline that first removes temporally redundant frames using adaptive clustering with Mahalanobis and Hamming dissimilarities and then segments the selected frames using a task-specific PSPNet adaptation with a ResNet-50 backbone, dilated convolution, cross-layer feature fusion, and channel attention. The contribution is an application-oriented integration and evaluation of established components rather than a new general-purpose segmentation architecture. The source pool contained 23,000 sports clips and approximately 1.47 million candidate frames; after quality control and redundancy reduction, 130,000 annotated frames were split by source video into 91,000 training, 26,000 validation, and 13,000 test frames. On the reported test protocol, the model obtained 88.93% mean Intersection over Union (mIoU), 95.36% pixel accuracy, 93.45% Dice, and 89.64% Cohen's kappa. The measured segmentation-model inference time was 0.18 s per selected 1024 x 576 frame on an RTX 3090; keyframe selection required 0.24-0.26 s per evaluated clip. These stage-specific values should not be interpreted as end-to-end real-time throughput. Because archived prediction masks and training logs were insufficient to reconstruct per-class, boundary-IoU, small-object, or contemporary-transformer baseline results, the present evidence supports only a proof-of-concept claim. Reproducible comparison with U-Net, DeepLabv3+, SegFormer, and Mask2Former remains required before a state-of-the-art claim can be made.