Autoencoder versus pre-trained CNN networks: deep-features applied to accelerate computationally expensive object detection in real-time video streams
Todd B. Howlett, Jeffrey T. Carlo, Vasanth Iyer, Bernard O. Abayowa, Alexander J. Aved · 2018
Traditional event detection from video frames are based on a batch or offline based algorithms: it is assumed that a single event is present within each video, and videos are processed, typically via a pre-processing algorithm which requires enormous amounts of computation and takes lots of CPU time to complete the task. While this can be suitable for tasks which have specified training and testing phases where time is not critical, it is entirely unacceptable for some real-world applications which require a prompt, real-time event interpretation on time. With the recent success of using multiple models for learning features such as generative adversarial autoencoder (GANS), we propose a two-model approach for real-time detection. Like GANs which learns the generative model of the dataset and further optimizes by using the discriminator which learn per sample difference between generated images. The proposed architecture uses a pre-trained model with a large dataset which is used to boost weekly labeled instances in parallel with deep-layers for the small aerial targets with a fraction of the computation time for training and detection with high accuracy. We emphasize previous work on unsupervised learning due to overheads in training labeled data in the sensor domain.