Comparative Analysis of VISEM Tracking Using YOLO, Vision Transformer, and Hybrid Model

Abstract

Automated sperm motility assessment requires reliable tracking of dense, visually similar cells in microscopy videos. This study compares three architectures on the VISEM-Tracking dataset: YOLO+IoU, Pure-ViT, and Hybrid YOLO+ViT. As for both ViT-based architectures, an additional 8×8 patch-size ablation was conducted to examine whether finer spatial tokenization reduces identity fragmentation. Tracking was evaluated using MOTA, IDF1, and identity switches, while downstream clinical consistency was measured using WHO-style motility distribution MAE. Results show that YOLO+IoU achieved the best tracking performance, with 51.49% MOTA, 58.16% IDF1, and 58.83 average identity switches. It also produced the lowest WHO distribution MAE at 14.68 percentage points. The 8×8 configuration reduced identity switches and total count error for the ViT-based trackers but failed to improve their WHO distribution MAE. Pure-ViT and Hybrid YOLO+ViT suffered from identity fragmentation because generic ViT embeddings could not reliably distinguish visually homogeneous sperm cells. These errors distorted trajectory-based motility classification. The findings indicate that spatial association remains more reliable than generic appearance-based matching for sperm tracking in high-frame-rate microscopic videos.