Energy-based Surpise Modeling using Pattern Theory



Egocentric perception has grown rapidly with the advent of immersive computing devices. Human gaze prediction is an important problem in analyzing egocentric videos and has largely been tackled through either saliency-based modeling or highly supervised learning. In this work, we tackle the problem of jointly predicting human gaze points and temporal segmentation of egocentric videos, in an unsupervised manner without using any training data. We introduce an \textit{unsupervised} computational model that draws inspiration from cognitive psychology models of human attention and event perception. We use Grenander's pattern theory formalism to represent spatial-temporal features and model \textit{surprise} as a mechanism to predict gaze fixation points and temporally segment egocentric videos. Extensive evaluation on two publicly available datasets - GTEA and GTEA+ datasets show that the proposed model is able to outperform all unsupervised baselines and some supervised gaze prediction baselines. Finally, we show that the model can also temporally segment egocentric videos with a performance comparable to more complex, fully supervised deep learning baselines.



Gaze Prediction Results



The predicted gaze point is illustrated as a gaze hatmap and the groundtruth gaze point is illustrated as the red point.

Event Segmentation Results



Left shows the segmentation prediction from the pattern theory framework in a streaming fashion. The top Gantt chart shows the groundtruth values and the bottom Gantt chart shows the predicted values. On the right, we present the quantitative evaluation on the GTEA dataset. The metric used is accuracy obtained after Hungarian matching.

Supervision Approach Accuracy
Full Spatial-CNN 54.1%
Full Bi-LSTM 55.5%
Full Dilated TCN 58.3%
Full ST-CNN 60.6%
Full TCN 64.1%
Full EgoNet + TDD 64.4%
None Our Approach 57.9%


Code, Paper and Extras

  • Find training/evaluation code on Github [COMING SOON!].
  • Find the paper here

Bibtex

@misc{aakur2020unsupervised,
    title={Unsupervised Gaze Prediction in Egocentric Videos by Energy-based Surprise Modeling},
    author={Sathyanarayanan N. Aakur and Arunkumar Bagavathi},
    year={2020},
    eprint={2001.11580},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}