Other Datasets
Note a lot of the datasets I've looked at are mainly images gathered from the web and Flikr that people took and posted so they might be of limited use for Blue Iris object detection. But if you have a limited set from your own cameras to work with they could improve that base set.
You might want to start with this Introduction to YOLOv5 Object Detection with Tutorial to learn more about model performance to decide which you want to use. Default is yolov5m (medium) which is probably best for most Blue Iris setups.
Kaggle is a good place to look for any other datasets you might want as many seem to list theirs there.
Roboflow another place where people list their datasets
abhineet123/Deep-Learning-for-Tracking-and-Detection Has a list of datasets by categories.
Actually links to multiple datasets. Huge number of images but they are from all over the world, many have no animal even in them and the mapping info is pretty much useless so filtering will probably manual to create a set for training. For more info see Kaggle YOLO V5 Animal Detection
520 images of 3 classes. Appears to be mainly close ups. Mapping files are in Pascal VOC / XML format which will need converting.
37 category pet dataset with roughly 200 images for each class. The images have a large variations in scale, pose and lighting. All images have an associated ground truth annotation of breed, head ROI, and pixel level trimap segmentation.
11,788 of 200 classes
github repo with info about dataset
2349 total bird images. But all the objects are mapped to the one class, bird, and mapping files are in json format that needs converting. See issue 3
4463 images in folders by state. The only mapping data is a csv file mapping file names to state. And the images are cropped close to the plates.
Only 350 images from all over the world so fairly generic but can download in many formats.
The set the DeepStack_OpenLogo model was trained with. Note pretrained model appears to be a yolov5x model
27,083 total images for 353 classes.
Has links to other datasets but some are gone or moved.
The dataset the dark model was trained with. Note pretrained model appears to be a yolov5l model
7,363 low-light images for 12 object classes but in folders by object type without mapping files.
See Convert And Merge my experience trying to merge some of these datasets
502 images for one class, fire. All types from people lighting cigarettes to house fires. A lot are vehicle files.
All the mapping files are in Pascal VOC / XML format which will need converting. See issue 5
Only 100 unmapped image sample of the full dataset which is available upon request. But looks useful for spotting small close fires or from a tower cam which is what I want.
14671 jpg and 1094 png images of smoke and fire that look perfect for my needs from the examples though there is no mapping data so will need mostly manually labeled. On closer examination it appears almost all of the smoke training images are sequential frame grabs with smoke plumes Photoshopped in. This is because it is for TensorFlow training, not YOLO.
755 Mainly outdoor-fire images and 244 non-fire images for computer vision tasks. Mainly large fires close up.
These looked possibly useful for fire detection from security cams.
Note one image in "non-fire" is of a fire. Most are forest shots. Many with people and a few with smoke.
736 tower cam images in labeled yolov5 format classes are
| ID | Class |
|---|---|
| 0 | Not a wildFire |
| 1 | WildFire |
All images of just smoke from fires so you probably would want to rename the "Not a wildFire" class "smoke" and drop class "Wildfire". Also the test images do not all pass a model trained on their training images. See Convert And Merge for more details.
Experimentally Defined Convolutional Neural Network Architecture Variants for Non-temporal Real-time Fire Detection
Not useful for Blue Iris + DeepStack training but an interesting read on where things are going with the use of super pixels.