Revisiting Risky Tackle Detection with Vision Transformers

Syed Ahsan Masud Zaidi; Lior Shamir; Scott Dietrich

cs.CVSep 28, 2026

Revisiting Risky Tackle Detection with Vision Transformers

Authors: Syed Ahsan Masud Zaidi, Lior Shamir, Scott Dietrich

Organizations: Kansas State University, Manhattan, KS, USA · Albright College, Reading, PA, USA

Abstract

This paper is a Track 2 reproducibility companion to an ICPR 2026 study on risky tackle detection in American football prac- tice videos. The original work fine-tuned a Video Vision Transformer (ViViT) on 733 clips labeled with the SATT-3 rubric. It used focal loss, Taguchi L18 augmentation, and 5-fold cross-validation. It reported risky- class recall of 0.67 and risky-class F1 of 0.59. This companion documents the released artifact and traces those numbers to specific scripts, fold out- puts, and aggregation files. The reproduced headline is run_15. It com- bines Gaussian noise with static brightness decrease and uses no rotation and no flip. Its fold-mean risky recall is 0.667 and its fold-mean risky F1 is 0.588. These values match the published headline after rounding. The ablation shows that brightness is the dominant factor. Its risky-recall main-effect range is 0.055, which is larger than the ranges for rotation, flip, and noise. Without augmentation, ViViT reaches risky recall of 0.545 and does not exceed the C3D baseline of 0.583. The raw clips show iden- tifiable student athletes, so they cannot be redistributed. The artifact provides a public sample for pipeline checks and a controlled route for full-data review.

Figures & tables

Paper item	Script or source	Config.	Output artifact
Headline and Table 3	vivit_train_taguchi.py plus consolidate_metrics.py	50 epochs	metrics_summary.csv ; consolidated heatmap
$L_{18}$ main effects in Table 5	heatmap values from opt_* metrics	5 folds	performance heatmap
Complete heatmap audit in Table 6	consolidate_metrics.py	fold means	consolidated_metrics.csv
Public-sample build	Taguchi_datasets.py	demo split	taguchi_runs/run_*/
Full-data fold directories	prepared taguchi_runs	5-fold $L_{18}$	run_/fold_/
Full-sweep job submission	SLURM launcher	20 runs, 5 folds	job logs

Table 1: Paper-to-code mapping for the released artifact.

Setting	Value
Training script	vivit_train_taguchi.py
Backbone	ViViT [ 1 ] ; google/vivit-b-16x2-kinetics400 [ 2 ]
Frames and resolution	32 frames at $224\times 224$
Temporal window	15 frames before FPOC and 16 frames after FPOC
Batch and precision	batch size 2; bf16 on supported GPUs
Gradient accumulation	8 steps; effective batch size 16

Table 2: Training configuration for the reproduced headline lineage.

Configuration	Risky recall	Risky F1	Risky prec.	Accuracy	Safe recall
C3D baseline [ 4 ]	0.583	0.560	0.538	0.711	0.769
ViViT, no augmentation	0.545	0.530	0.522	0.660	0.724
ViViT, oversample only	0.537	0.540	0.547	0.679	0.757
Best Taguchi ( run_15 )	0.667	0.588	0.535	0.669	0.670
Fig. 4 heatmap in [ 8 ]	0.667	0.588	0.535	0.669	0.670
Mean $\pm$ std, runs 01 to 18	$0.572\pm 0.041$	$0.550\pm 0.020$	$0.541\pm 0.027$	$0.672\pm 0.016$	$0.726\pm 0.037$

Table 3: Baseline comparison and headline result. Values are 5-fold means from the threshold-tuned heatmap.

Factor	Code	Levels	Count
Noise	$A$	None; Gaussian noise	2
Brightness	$B$	Static increase; static decrease; same	3
Rotate	$C$	Left; right; none	3
Flip	$D$	Horizontal; vertical; none	3
Full factorial space			54
$L_{18}$ runs			18

Table 4: Taguchi

L_{18}

factors for the reported 733-clip heatmap. Brightness uses static increase and static decrease levels.

Factor	Level 1	Level 2	Level 3	Range	Best
Noise ( $A$ )	None: 0.564	Noise: 0.580	–	0.016	Noise
Brightness ( $B$ )	Incr.: 0.552	Decr.: 0.607	Same: 0.559	0.055	Decrease
Rotate ( $C$ )	Left: 0.558	Right: 0.574	None: 0.586	0.028	None
Flip ( $D$ )	Horiz.: 0.582	Vert.: 0.571	None: 0.565	0.017	Horizontal

Table 5: Taguchi main effects for risky-class recall. Values are computed from the 18 threshold-tuned heatmap recalls.

Configuration	Accuracy	Risky prec.	Risky recall	Risky F1	Safe recall
C3D baseline	0.711	0.538	0.583	0.560	0.769
No augmentation	0.660	0.522	0.545	0.530	0.724
Oversampled no augmentation	0.679	0.547	0.537	0.540	0.757
run_01	0.674	0.536	0.560	0.544	0.736
run_02	0.673	0.552	0.498	0.521	0.768
run_03	0.670	0.535	0.563	0.542	0.728

Table 6: Complete heatmap values used for the reproduction check. Values are fold means.

Explore similar work

CheckOne: Lightweight Fault Detection and Mitigation for Vision Transformers
Aug 3, 2026Mohammad Hasan Ahmadilivani, Sven-Markus Loorits, Jaan RaikSelf-Supervised Vision Transformers Vision Transformer
Tallinn University of Technology, Tallinn, Estonia
EviViT: Evidence-Adaptive Vision Transformers for Fine-Grained Perception
Sep 29, 2026Yaoxin Niu, Zhangquan Chen, Yang Zhang +5Self-Supervised Vision Transformers Fine-Grained Perception
Tsinghua University · Peng Cheng Laboratory · The Hong Kong University of Science and Technology +3
An Open-Source Two-Stage Computer Vision Pipeline for Fine-Grained Vehicle Classification using Vision Transformers
Jun 3, 2026Gandhimathi Padmanaban, Fred FengAutomotive Detection Computer Vision
Department of Industrial Manufacturing and Systems Engineering, University of Michigan-Dearborn, Michigan, USA, 48128

Aug 3, 2026, cs.ARJ/K move · Enter open · S save

CheckOne: Lightweight Fault Detection and Mitigation for Vision Transformers

Mohammad Hasan Ahmadilivani, Sven-Markus Loorits, Jaan Raik

Tallinn University of Technology, Tallinn, Estonia

Self-Supervised Vision Transformers Vision Transformer Fault Diagnosis Fault Tolerance Mitigation Symmetry

Open paper AskPDF arXiv

The wide adoption of Vision Transformers (ViTs) in safety-critical applications raises reliability concerns related to hardware faults. Algorithm-Based Fault Tolerance (ABFT) methods have emerged as lightweight and symmetric protection mechanisms for DNNs. However, they are particularly challenging for ViTs due to their significant computational requirements. This work comprehensively evaluates the reliability of ViTs, emphasizing the need for symmetric protection in their layers. Furthermore, we present CheckOne, a novel, cost-effective method for fault detection and mitigation in ViTs that significantly reduces the computational cost compared to conventional ABFT. Through extensive experiments with multiple ViTs, CheckOne mitigates critical faults by up to

26\times

and achieves an average 3.8x higher performance than ABFT in ViTs.

Abstract

Explore similar work

CheckOne: Lightweight Fault Detection and Mitigation for Vision Transformers

EviViT: Evidence-Adaptive Vision Transformers for Fine-Grained Perception

An Open-Source Two-Stage Computer Vision Pipeline for Fine-Grained Vehicle Classification using Vision Transformers