PaperlayerView exact revision on arXiv

Source-linked AI summary

Explaining Basketball Shot Classifications with Grad-CAM: An Investigation of Shortcut Reliance with Spatiotemporal Correlations

Hongyi Tian, Robin Cohen

arXiv:2610.33263v1cs.CV

TL;DR

Evidence used by basketball shot-classification models is poorly understood, and explainability in basketball action recognition remains under-analyzed. Grad-CAM faithfulness was evaluated on 1,121 testing clips by replacing an additional 10% of highest-ranked spatiotemporal locations per step; 20 clips were reviewed to define Gaussian-blurred and random-control regions. 43.7% accuracy after central blurring was below random-blurred accuracy of 50.9% and original test accuracy of 69.3%.

  • Problem

    Evidence used by basketball shot-classification models is poorly understood, and explainability in basketball action recognition remains under-analyzed.

    Source: Motivation · p. 1

  • Method

    Grad-CAM faithfulness was evaluated on 1,121 testing clips by replacing an additional 10% of highest-ranked spatiotemporal locations per step; 20 clips were reviewed to define Gaussian-blurred and random-control regions.

    Source: Explanation and Intervention Protocol · pp. 1–2

  • Results

    43.7% accuracy after central blurring was below random-blurred accuracy of 50.9% and original test accuracy of 69.3%.

    Source: Results and Limitations · p. 2

  • Takeaways & Limitations

    The study offers explanation workflows adaptable to other CNN-based action-recognition models and audits of shortcut learning aimed at reducing unfair athlete penalization.

    Source: Discussion and Future Work · p. 2

    Cross-architecture comparisons of shortcut learning and training with shortcut-mitigation efforts remain future work.

    Source: Discussion and Future Work · p. 2

Abstract

from arXiv · show

Basketball action recognition is used in high-stakes scenarios to inform such tasks as specialized coaching, yet the evidence used by its deep-learning models is poorly understood. This paper investigates whether there is evidence that an R(2+1)D-18 model trained on the Basketball-51 dataset relies on spurious spatiotemporal correlations when classifying basketball shot outcomes. Gradient-weighted Class Activation Mapping (Grad-CAM) was used to generate spatiotemporal saliency maps, whose faithfulness was evaluated by progressively blurring the most salient pixels. Twenty saliency maps were then visually reviewed for recurring salient regions. Grad-CAM highlighted spatially central regions of clips including the shooter, net, and occasionally the courtside crowd. Blurring interventions were conducted to support these patterns. Blurring the central region resulted in an accuracy of 43.7%, compared to the original test accuracy of 69.3% and random-blurred accuracy of 50.9%. The additional 7.2 percentage-point reduction from central blurring compared to random blurring suggests that the model relies on task-relevant visual regions; this highlights the need for continual audits to ensure trustworthy, equitable sports analytics.

Motivation

The evidence used by basketball shot-classification models is poorly understood, and explainability in basketball action recognition remains under-analyzed. This study examines whether a 3D-CNN relies on spurious correlations rather than authentic shooting mechanics.

Motivation

  • Shortcut learning can exploit unintended cues, such as scoreboard changes, instead of task-relevant evidence like shooting posture and net movement.Changing scoreboards and crowd sizes across arenas and leagues may impair real-life performance and risk unfair or inaccurate athlete evaluation.Source: Motivation · p. 1
  • The study investigates whether a 3D-CNN uses spurious spatiotemporal correlations instead of authentic mechanics to classify shot outcomes.The paper identifies explainability in basketball action recognition as under-analyzed.Source: Motivation · p. 1
  • The proposed workflow combines action recognition, localized spatiotemporal explanations, and sports analytics to distinguish task-relevant cues from spurious correlations.Grad-CAM produces class-discriminative localization maps that highlight video regions salient to a model’s decision.Source: Motivation · p. 1

Data and Model

The study uses Basketball-51 clips from disjoint games and fine-tunes a pretrained R(2+1)D-18 model for eight shot classes.

Data and Model

  • Basketball-51 supplies shooting clips from 51 broadcast NBA games, divided into eight classes, with 7,750 training, 1,440 validation, and 1,121 test clips.The respective splits draw from 39, 7, and 5 disjoint games.Source: Data and Model · p. 1
  • The baseline is a Kinetics-400-pretrained R(2+1)D-18 fine-tuned on 16-frame, 112×112-pixel clips.Its spatiotemporal convolutions factor each 3D convolution into a 2D spatial convolution followed by a 1D temporal convolution.Source: Data and Model · p. 1
  • The replacement classification head maps 512 pooled features to eight class scores, and the checkpoint with the highest validation accuracy is selected.Training stops after validation accuracy fails to improve for two consecutive epochs.Source: Data and Model · p. 1

Explanation and Intervention Protocol

The protocol tests Grad-CAM saliency by progressively blurring highly ranked spatiotemporal locations and comparing them with random-location controls. It also examines recurring regions in reviewed clips through targeted and control blurring.

Explanation and Intervention Protocol

  • At each perturbation step, an additional 10% of the highest-ranked spatiotemporal locations is replaced with Gaussian-blurred frame values.The corresponding frame blur uses a 7×7 Gaussian kernel.Source: Explanation and Intervention Protocol · pp. 1–2
  • A random ordering of locations serves as a control, with probabilities averaged across all 1,121 test clips at each perturbation level.Source: Explanation and Intervention Protocol · pp. 1–2
  • After visually reviewing saliency maps from 20 test clips, the protocol blurs recurring salient regions in the test set and compares accuracy with equal-sized random-region blurring.The stated purpose is to help determine whether shortcut learning is occurring.Source: Explanation and Intervention Protocol · p. 2

Results and Limitations

Targeted central blurring produces a larger accuracy drop than random blurring, while Grad-CAM perturbations reduce confidence more when they target the most salient pixels. Visual review finds stable, shot-dependent salient regions, with some crowd saliency in three-point clips.

Results and Limitations

  • The retained epoch-4 checkpoint reached 64.4% peak validation accuracy and 1.267 validation cross-entropy loss; training stopped after epoch 6.Validation accuracy had failed to improve for two consecutive epochs at stopping.Source: Results and Limitations · p. 2
  • Blurring the most salient pixels causes an immediate, substantial drop in average confidence, while random-pixel blurring causes a much smaller drop.Figure 1 reports the average softmax probability for the original predicted label across 1,121 cases.Source: Results and Limitations · p. 2, Figure 1 · p. 2
  • Across 20 reviewed clips, the most salient regions remain largely stable over time, with peak-saliency frames between ball release and its fall to the ground.Figure 2 shows a representative Grad-CAM sequence.Source: Results and Limitations · p. 2, Figure 2 · p. 2
  • 43.7% accuracy after central blurring is 7.2 percentage points below random-blurred accuracy of 50.9%, versus 69.3% baseline accuracy.The targeted blur covers the centre 60×60 pixel region; the authors say this estimates the effect because it does not separate player, ball, net, or background.Source: Results and Limitations · p. 2

Discussion and Future Work

The discussion presents adaptable XAI workflows and shortcut auditing as contributions for sports analytics and spatiotemporal models. Future work will compare architectures and examine masking potential spurious cues.

Related papers

Loading 2610.33263v1…