Deep Tracks: Using CNNs and Procedurally Simulated Data for Automated Vertebrate Footprint Classification

Simulated dataset We provide the simulated data (.png) used to train neural network classifiers, separated by their corresponding labels and group (train or validation). The dataset contains a total of 40,000 simulated footprints images, belonging to ten different vertebrate classes. The footprints belong to the following clades: Mammalia (coyotes, foxes, bears, otters, squirrels, raccoons, deer), avian Dinosauria (turkeys) and non-avian Dinosauria (theropods, sauropods). The images were created by Deep Tracks (DOI: 10.5281/ZENODO.15092442). Code We provide the code used by the best model configuration obtained. The first is the Python code used to pre-train the Convolutional Neural Network (CNN) on the simulated dataset, the second is the Python code to finetune the previously trained model to real footprint photographs. Trained model weights and information We provide the trained models information, (e.g., model weights, opt, The trained models information for both real and simulated datasets are avaia Files provided Simulated_data.zip - contains the images of the simuated footprints(.png) BestSimulatedModel.zip - contains the trained model information for the simulated data (.pth) BestRealModel.zip - contains the trained model information for the footprint photographs (.pth) model_train_pretrain_simulateddata_efficientnet_b0.py - contains the code to pre-tain the model (.py) model_train_finetune_to_realdata_efficientnet_b0.py - contains the code to finetune the pre-trained model on a real-world dataset (.py)