The following paper is a follow-up from the deep learning course at the TU Delft with Prof J. van Gemert. There was a project to be made, this had some success, and after some hard work due the summer holiday, it was accepted by the NCCV 2019. The paper is called Fine-grained Classification of Rowing teams, because we have to look at a few pixels (fine-gradient), of the image that determines which rowing association that team belongs. In our paper, we have shown that this was actually quite an easy task for the neural networks. In the following picture, we show.
Above an example is shown from our rowing boat dataset. Here a crew from Proteus-Eretes and Njord is visible. Under specific lighting conditions, the different shades of blue can appear similar. The similarities will make it difficult for a deep learning network to classify the teams correctly.
Abstract
Fine-grained classification tasks such as identifying different breeds of dog are quite challenging as visual differences
between categories is quite small and can be easily overwhelmed by external factors such as object pose, lighting,
etc. This work focuses on the specific case of classifying
rowing teams from various associations.
Currently, the photos are taken at rowing competitions and
are manually classified by a small set of members, in what
is a painstaking process. To alleviate this, Deep learning
models can be utilised as a faster method to classify the images. Recent studies show that localising the manually defined parts, and modelling based on these parts, improves
on vanilla convolution models, so this work also investigates the detection of clothing attributes.
The networks were trained and tested on a partially labelled
data set mainly consisting of rowers from multiple associations.
This paper resulted in the classification of up to ten rowing
associations by using deep learning networks the smaller
VGG network achieved 90.1% accuracy whereas ResNet
was limited to 87.20%. Adding attention to the ResNet
resulted into a drop of performance as only 78.10% was
achieved.
Paper submission link: https://arxiv.org/abs/1912.05393
Paper source code link: github.com/martijnvwezel/ResNet-Attention-layer-custom-data-set