A set of datasets such as UCF101-Action Recognition Dataset (UCF101-ARD), FCVID and Hollywood2 were intially discovered and downloaded. Out of those the UCF101-ARD was selected because of several properties such as high human figure density, different kinds of action frequencies, many landscape variations like sea, beach, sky river, etc. and many context changes.
The first required step of the project is preparing a dataset that the model can be trained on. For that, at the initial phase, “Hollywood 2 - Human action and scene dataset” was selected. This dataset has only around 2500 videos annotated with different action tags. Only the video part from the dataset was used. Since the video colorization, is a complex problem, with this small dataset, the model started to overfit.
Hence in the second stage, it was decided to use the FCVID dataset. The primary use case of FCVID dataset is video content detection and action detection. It contains 91,233 videos and 239 categories, where 183 are events and 56 are objects, scenes, etc., taken from the web with a total duration of 4,232 hours with an average video duration of 167 seconds. The dataset was annotated manually with 239 different classes. Only the videos were required for the experiment. Around 50000 videos were randomly selected from the dataset and processed for training.
Thus, in the FlowChroma, around 50,000 preprocessed five frame videos from the FCVID video dataset were used to train the model. Hollywood2 actions dataset containing around 2500 videos was also processed to train Proof of Concepts.
Associated Publication: -
Paper Title: FlowChroma - A Deep Recurrent Neural Network for Video Colorization
Published in: International Conference on Image Analysis and Recognition
ICIAR 2020: Image Analysis and Recognition pp 16-29
DOI: https://doi.org/10.1007/978-3-030-50347-5_2
Citation: -
Wijesinghe T., Abeysinghe C., Wijayakoon C., Jayathilake L., Thayasivam U. (2020) FlowChroma - A Deep Recurrent Neural Network for Video Colorization. In: Campilho A., Karray F., Wang Z. (eds) Image Analysis and Recognition. ICIAR 2020. Lecture Notes in Computer Science, vol 12131. Springer, Cham. https://doi.org/10.1007/978-3-030-50347-5_2
