Create README.md

README.md

@@ -0,0 +1,109 @@
+---
+license: mit
+language:
+- en
+pretty_name: Lanternfly Image Classifier Training Dataset
+---
+
+
+# Dataset Card for {{ pretty_name | default("Dataset Name", true) }}
+
+This dataset is the training dataset for 24-679 Project 1: Lanternfly Tracker
+It is composed of 360 original lanternfly photos, 150 original photos with no lanternflies, and 800 original photos
+from nature, urban, and other insect datasets listed below. 
+
+These were augmented 50X to 65.1k augmented images. 
+
+
+## Dataset Details
+
+### Dataset Description
+
+- **Curated by:** Carnegie Mellon University: 24-679
+- **Shared by [optional]:** Devin DeCosmo
+- **Language(s) (NLP):** English
+- **License:** MIT
+
+### Dataset Sources [optional]
+
+Original Lanternfly Datasets
+  rlogh/lanternfly-data: Original Lanternfly Dataset, 229 unmarked photos
+  rlogh/lanternfly_swatter_training: Dataset with geolocal data: 165 photos
+  
+Original Negative Datasets:
+  rlogh/negativesirl: Negatives dataset, images of outdoor environements and people with no lanternflies. 107 photos
+
+
+Total: 501 original images
+
+Imported Datasets
+
+
+
+## Uses
+
+These images were used to train the EfficientNetB1 model, ddecosmo/lanternfly_classifier, on how to classify images 
+as containing or not containing lanternflies. 
+
+
+### Direct Use
+
+The direct use is identifying photographs containing lanterflies so this could be used for tracking purposes. 
+
+### Out-of-Scope Use
+
+In future, this model could be adapted to identify other types of insect within this dataset. 
+
+
+## Dataset Structure
+
+This dataset consists of two splits
+An original split with 1.3k photos
+An artificial split with 65.1k photos
+
+The tasks fall into 3 categories based on the building pictured
+1. Lanternflies, all original photos
+2. Other Insect, all 3rd party datasets
+3. No insect, original photos and 3rd party datasets
+
+## Dataset Creation
+
+### Source Data
+
+This data is sourced by the creators, Devin and Rumi for all original photos
+
+Additional datasets can be found here,
+uoft-cs/cifar100
+AI-Lab-Makerere/beans
+Francesco/insects-mytwu
+
+#### Data Collection and Processing
+
+Original datasets were collected using the mobile phones of the authors. 
+
+Additional datasets were recommended by Gemini AI and then validated as fitting the purpose, type, and scope of this process.
+uoft-cs/cifar100: This is a general image identifier with no insect class. Used for no insect for generalizability
+AI-Lab-Makerere/beans: This dataset is focused on vegetation with and without disease, this is used to train the model to recognize 
+vegetation without insects/lanterflies.
+Francesco/insects-mytwu: This is an object detection dataset used for identifying insects as subjects, not including lanterflies. 
+We are using it train a seperate non-lanternfly insect class. 
+
+#### Who are the source data producers?
+
+Original data was produced by the authors. 
+
+Additional datasets were produced by, 
+uoft-cs/cifar100: Created by University of Toronto Computer Science
+AI-Lab-Makerere/beans: Created by AI Lab Makere
+Francesco/insects-mytwu: Created by Fanscesco Sovrano
+
+## Bias, Risks, and Limitations
+
+The main risk of this dataset is the lanternfly split. It contains only images of singular lanternflies on the ground.
+Normally on concrete or asphalt. This severly limits the scope of the environments these creatures appear in.
+Incorporating blob detection or YOLO into future models could mitigate this by focusing on the subject. 
+
+### Recommendations
+
+This is a large dataset, and has been shown to accurately classify lanternflies, but there are many edge cases when it does not work correctly. 
+In order to take this into account, using new types of models with subject detection can make use of the many images while improving model accuracy.