MegaDetector Object Detection: performance on high noise animal camera trap data

Mega-Detector is a prestigious Microsoft product for object detection [1], well documented, and used by many animal-analytics (and other applications) companies [2]. Its latest version “MDv5” is even pre-trained on Animal CameraTrap data. There are multiple data challenges regarding Animal Camera Trap data like Illumination, Motion Blur, Small region of interest, Occlusion, Camouflage, Perspective, Weather Conditions, Camera Malfunctions, Temporal Changes, and Non-Animal Variability [3]. In a previous post [4] a method was discussed to extract images (see below the Noise Figures Extracted section) with these challenges from iNaturalist 2021 data (or any other animal camera trap data). The method does not classify noise images into the mentioned data challenges categories though. That was done by inspecting the extracted images and guessing what possible data challenge category could it be in. In this post, we discuss the performance of MegaDetector on the extracted images. The code base utilized is the google-collab notebook tutorial published on the repo itself [5]. Looking at the 2- Mega Detector Output section in this post, it seems like MegaDetector was able to put very accurate boundary boxes around animals. Except with the ones with Camouflage and a small region of interest combined. It was surprising to see it beat Occlusion specifically. In one of the images, MegaDetector thought that an animal-shaped rock was also an animal but only with a 34% probability of it being one. Finally, images that had no animals were just left blank, it did not draw any boxes in them, which is good, those can even be removed completely.

All in all, the performance of MegaDetector was really good in detecting and putting boundary boxes on animals even with noisy images.

1- Noise Figures Extracted

Camouflage and small region of interest

Illumination

Motion blur

Occlusion

small region of interest

Occlusion

Perspective

Motion Blurr and Perspective

No Animal

No Animal

No Animal

2- Mega Detector Output

Refernces

[1] https://github.com/microsoft/CameraTraps/blob/main/megadetector.md

[2]https://github.com/microsoft/CameraTraps/#who-is-using-megadetector

[3] https://github.com/visipedia/iwildcam_comp

[4] https://medium.com/@emad-ezzeldin4/debugging-computer-vision-image-classification-removing-noisy-images-2fb7c518930b

[5] https://colab.research.google.com/github/microsoft/CameraTraps/blob/master/detection/megadetector_colab.ipynb#scrollTo=vUXNQZtwEYiQ