Home

Awesome

Elephant Detection

Using Keras-RetinaNet to detect elephants from aerial images

Flatiron School DC-DS-08262019 Capstone Project

In this project I wanted to explore the realm of computer vision, and jumped straight into the deep pool of object detection.

Data

Source: Naudé, Johannes J., & Joubert, Deon. (2019). The Aerial Elephant Dataset. Zenodo. http://doi.org/10.5281/zenodo.3234780

From the paper:

(A) challenging dataset to enable research on game detection under real-world conditions. The dataset consists of 2101 images containing a total of 15 511 African bush elephants in their natural habitats, imaged with a consistent methodology over a range of background types, resolutions and times-of-day.

Data Image Example

Methods and Libraries

I used the Keras implementation of RetinaNet by fizyr, which is based on Focal Loss for Dense Object Detection Paper by Tsung-Yi Lin, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar.

As a supplement, this anchor optimization tool by @martinzlocha is also used to make sure the right configuration (anchor ratios and scales) is implemented. This is especially important due to the (extremely) small object sizes to be detected.

This is a single class object detection task (class 0: elephant)

Work Environment

Project Steps and Results

Model performance at this point:

Inference Example 1 Inference Example 2 Inference Example 3

Discussion and Future Work

The model resulted in a lot of false positives (about 7 false positives for every true positive), and it does a better job (higher recall, precision) on some landscapes than others. To improve performance, the following steps can be taken: