@article{,
title= {N+1 fish, N+2 fish dataset (train_videos)},
journal= {},
author= {N+1 fish, N+2 fish},
year= {},
url= {https://www.drivendata.org/competitions/48/identify-fish-challenge},
abstract= {Our video data was collected for an ongoing monitoring project involving our partners at The Nature Conservancy-Massachusetts (TNC) and the Gulf of Maine Research Institute (GMRI). We worked with the fishermen to create a dataset of video footage from different boats that can be released to the public. 

The video data is in the form of standard and well supported MP4 v2 format with H.264 compression. These videos are on average 35 minutes long at a resolution of 640x480 and and vary between 150MB to 550MB in size. We will also provide cropped fish images to aid in training complementary models. Our annotations will be categorial with each fish having a single species. 

There are six species of interest for this competition, which appear in a non-balanced proportion. As we aim to achieve performance on each of these species, we will be compiling a data set that is mostly balanced across these species. Complicating factors include different catch distributions across different vessels, which can have an adverse performance on the types of algorithms submitted if people attempt to game the system by parameterizing these distributions. We are attempting to balance the data set by providing at least 100 unique fish for each species within the videos from multiple vessels.




![](https://i.imgur.com/qHK6WUz.png)

![](https://i.imgur.com/T7UDAJA.png)},
keywords= {},
terms= {},
license= {},
superseded= {}
}