Context This project was initiated as a test of the Nvidia DIGITS and NVCaffe software and attempted to assess what sort of images are required for a good dataset. I managed an mAP of 20% using the BVLC_googlenet.cafemodel and was able to successfully infer wasps using the Nvidia Jetson TX2. There's a video here: [WASP IN A JAR][1]. The main project page, with instructions on how to use DIGITS and Amazon Web Services is here: [MAIN PROJECT PAGE][2] Content There are about 3,000 images of the wasps, each labelled with boundary boxes for object inference. Some of the images are partially occluded and truncated, but the labels DO NOT describe these attributes. Acknowledgements We wouldn't be here without the help of others. If you owe any attributions or thanks, include them here along with any citations of past research. Inspiration Please reply with some criticism of the images. Some of my questions are: 1. Should all the images be pristine quality? 2. Should the images reflect 'real life' camera environments? 3. What % of the images should be blurred / out of focus? 4. How important is it to label the occlusions and truncations? 5. How important is having different backgrounds in the images? ….. Thanks! [1]: https://www.youtube.com/watch?v=01MJmSwYuIo [2]: https://hackaday.io/project/161581-wasp-and-asian-hornet-sentry-gun