Recent searches

No recent searches

Search options

Only available when logged in.
ecoevo.social is one of the many independent Mastodon servers you can use to participate in the fediverse.
Dedicated to Ecology and Evolution. We welcome academics, students, industry scientists, folks from other fields with links to E&E, scientific societies, and nature enthusiasts in general.

Administered by:

Server stats:

535
active users

ecoevo.social: AboutProfiles directoryPrivacy policy

Mastodon: AboutGet the appKeyboard shortcutsView source codev4.4.3

Sarah Dal

In a very boring afternoon, I counted by hand about 1,900 seals in some drone imagery. With a possibility of flying a colony that might have 30,000 seals, I thought I should probably look into automatic counting with and plugin..

WAID is a dataset containing images of Sheep, Cattle, Seal, Camelus, Zebra, or Kiang. It's huge (14,375 images) and I just needed Seals, so I wrote a quick python script to remove images that had no seals from the dataset.
mdpi.com/2076-3417/13/18/10397

MDPIWAID: A Large-Scale Dataset for Wildlife Detection with DronesDrones are widely used for wildlife monitoring. Deep learning algorithms are key to the success of monitoring wildlife with drones, although they face the problem of detecting small targets. To solve this problem, we have introduced the SE-YOLO model, which incorporates a channel self-attention mechanism into the advanced real-time object detection algorithm YOLOv7, enabling the model to perform effectively on small targets. However, there is another barrier; the lack of publicly available UAV wildlife aerial datasets hampers research on UAV wildlife monitoring algorithms. To fill this gap, we present a large-scale, multi-class, high-quality dataset called WAID (Wildlife Aerial Images from Drone), which contains 14,375 UAV aerial images from different environmental conditions, covering six wildlife species and multiple habitat types. We conducted a statistical analysis experiment, an algorithm detection comparison experiment, and a dataset generalization experiment. The statistical analysis experiment demonstrated the dataset characteristics both quantitatively and intuitively. The comparison and generalization experiments compared different types of advanced algorithms as well as the SE-YOLO method from the perspective of the practical application of UAVs for wildlife monitoring. The experimental results show that WAID is suitable for the study of wildlife monitoring algorithms for UAVs, and SE-YOLO is the most effective in this scenario, with a mAP of up to 0.983. This study brings new methods, data, and inspiration to the field of wildlife monitoring by UAVs.
··Web
2boosts·5favorites
*
Sarah Dal

Frustratingly, it didn't do a great job with my imagery; I may have to sit for a few more hours tagging images and adding to the dataset, or the issue might be with the model?

Aerial shot of around 100 seals on a rocky shore. About a third have green bounding boxes where the Deepness has identified them correctly
Sarah Dal

Spent an hour tagging a few hundred more images of seals to the dataset and got a much better result! Thanks all who commented with ideas!

aerial view of grey seals on a rocky shoreline. The seals have purple boxes around them where the detection model has correctly detected a seal
#QGIS#Deepness#YOLO
David Lawrence Miller

@sarahdalgulls unsolicited observation (feel free to ignore): the majority of seals detected (I assume those are the green boxes) are in the same orientation (bottom left corner to top right corner diagonal. I remember going to a talk about something like this where they bulked-up the training set by taking the detections and rotating them at random and this improved performance.

(With apologies if you already knew about this! This is not really my area but in case it's helpful!)

Sarah Dal

@millerdl ahhh excellent catch thank you! I know about rotating the images when making the dataset, but as this was an already prepared dataset I didn't consider it.

I'm going to add to this one by tagging a couple hundred (yawn) seals from my pictures too, and see how that goes too

*
Raúl Nanclares 🍜

@sarahdalgulls
Data augmentation does rotations and other transformations without needing to add more images to the training dataset. If the ground has a different color it may have a huge impact on model performance.
@millerdl

Sarah Dal

@rnanclares @millerdl I think this is it, the training data were all of seals on sandy beaches, not rocky coast. I will train up my images as a dataset for future use

Raúl Nanclares 🍜

@sarahdalgulls
Why don't you try with the whole dataset? Not only training on seals may improve the results.
@millerdl

Raúl Nanclares 🍜

@sarahdalgulls @millerdl And then you can use your first model to help you create a new training dataset. Most of the time models require a "human in the loop" approach which means iterating and improving the training dataset.

Sarah Dal

@rnanclares I'm a very basic user, relying on non-programming tools (following this tutorial github.com/PUTvision/qgis-plug) Really just trying to see what's achievable on limited tech skills and no budget - all advice is greatly appreciated. I've added a bunch of my own seals to the training dataset and will run that now..

GitHubGitHub - PUTvision/qgis-plugin-deepness at masterDeepness is a remote sensing plugin that enables deep learning inference in QGIS - GitHub - PUTvision/qgis-plugin-deepness at master
Raúl Nanclares 🍜

@sarahdalgulls
Haven't tried the plugin. Adding your own data will surely improve the detection a lot.

David Lawrence Miller

@sarahdalgulls I remember talking to some folks at SMRU in St Andrews years ago when it was all by-eye counting and it sounded like an absolutely horrendous job!

TrendingLive feeds
About