Unsupervised Temporal Feature Aggregation for Event Detection in Unstructured Sports Videos

Abstract

Image-based sports analytics enable automatic retrieval of key events in a game to speed up the analytics process for human experts. However, most existing methods focus on structured television broadcast video datasets with a straight and fixed camera having minimum variability in the capturing pose. In this paper, we study the case of event detection in sports videos for unstructured environments with arbitrary camera angles. The transition from structured to unstructured video analysis produces multiple challenges that we address in our paper. Specifically, we identify and solve two major problems: unsupervised identification of players in an unstructured setting and generalization of the trained models to pose variations due to arbitrary shooting angles. For the first problem, we propose a temporal feature aggregation algorithm using person re-identification features to obtain high player retrieval precision by boosting a weak heuristic scoring method. Additionally, we propose a data augmentation technique, based on multi-modal image translation model, to reduce bias in the appearance of training samples. Experimental evaluations show that our proposed method improves precision for player retrieval from 0.78 to 0.86 for obliquely angled videos. Additionally, we obtain an improvement in F1 score for rally detection in table tennis videos from 0.79 in case of global frame-level features to 0.89 using our proposed player-level features. Please see the supplementary video submission at https://ibm.biz/BdzeZA.

A Comprehensive Review of Computer Vision in Sports: Open Issues, Future Trends and Research Directions

“…In sports such as table tennis [200], squash [201,202], and golf [203], ball detection and tracking and player pose detection [204] are challenging tasks. In ball-centric sports such as rugby, American football, handball, baseball, ball/player detection [205][206][207][208][209][210][211] and tracking [212][213][214][215][216][217][218][219][220][221], analyzing the action of the player [23,[222][223][224][225][226][227], event detection and classification [228][229][230][231][232], performance analysis of player [233][234][235], referee identification and gesture recognition are the major challenging tasks. Video highlight generation is a subclass of video summarization [236][237][238][239] which may be viewed as a subclass of sports video analysis.…”

“…It can be extended by incorporating artificial intelligence techniques. [228] Event detection in sports videos for unstructured environments with arbitrary camera angles.…”

A Novel Feature Aggregation Approach for Image Retrieval Using Local and Global Features

“…Recall = TP/(TP + FN) (10) where TP (True Positive) represents the correct classification of positive examples as positive examples; FN (False Negative) means to misclassify a positive example as a negative one; FP (false positive) means that negative cases are misclassified as positive ones.…”

“…However, Liang et al [8] and Liu et al [9] proved that the image features extracted from the last convolutional layer, namely the deep convolution features, were more significant than the fully-connected layer. Later, CNN based image retrieval models used deep convolution features and achieved good retrieval results, such as regional maximum activation of convolutions (R-MAC) [10], sparse bundle adjustment (SBA) [11].…”