Using DeepMeerkat for Fish Detection in Finland
I often get requests for training new biodiversity detection models for DeepMeerkat. While I continue to work on streamlining the training process, most users usually require some help along the way. A few weeks ago I was contacted by Viivi Kaasonen from Finland who had a particularly nasty challenge of detecting fish moving through a fishway. Some really tough conditions!
Viivi was able to use DeepMeerkat's training mode to scrape together 1300 images of fish and no fish. Here are some positive training examples.
I set aside an hour to retrain a model using 50 images withheld for model testing. Normally I'd suggest atleast 500 images for testing, but in ecology we often need every last data for training.
Here's the tensorboard result!
92% accuracy in identifying the testing data. That's pretty good!
Okay, some words of caution. 50 testing images is very small, and 1300 training clips is very small. So only time will tell how well these images represent the range of conditions encountered in Viivi's videos. Now all she needs to do is plug that model into DeepMeerkat.To set the new model, go to advanced settings and specify the new path to the model/ directory.
Okay, some words of caution. 50 testing images is very small, and 1300 training clips is very small. So only time will tell how well these images represent the range of conditions encountered in Viivi's videos. Now all she needs to do is plug that model into DeepMeerkat.To set the new model, go to advanced settings and specify the new path to the model/ directory.
Okay, that's all for now, in what hopefully is another win for biodiversity automation.