Skip to content

Home

Jump to bottom
Luke Duncan edited this page May 20, 2017 · 4 revisions

Welcome to the tfimgtools wiki!

Current Status

The codebase works, but has a lot of rough edges.

If you're a developer you can easily clone the repo and get busy, or you can simply install via pip. If you're an end-user, and will to jump through hoops, you can make this work on your environment. There are a lot of hoops to jump through though if you're not an engineer or have help from one.

Goal

  1. Provide a command line tool that non-engineers can use to classify images given they already have a model. The motivating usecase is to help zoologist who have a ton of pictures from camera traps who need to classify if the image is of an animal or not. Think of this as a step towards automating Snapshot Serengeti. It's a similar use-case for friends of the author.

  2. Be an excuse for the author(s) to play with tensor flow with a meaningful deliverable

  3. Build up a base of generic tools for image classification that are hopefully useful to the author(s) and maybe others.

Needed work (Help Wanted)

  • Packaging and Distribution for end-users -- A point and click installation on Windows or Mac that doesn't require the user to know how to install python and (in the case of Windows). See Goal 1.

  • Code reviews welcome. The author hasn't worked extensively in Python. Help his code become more idiomatic!

  • I imagine a CLI to be developed for emitting image classification statistics for comparative analysis between multiple models on use-case specific data sets (non-benchmark).

  • I imagine a CLI to be developed for performing classifications against benchmark datasets (mnist, ImageNet, etc) for comparative analysis. This may or may not need to include training.

  • If you have access to the Snapshot Serngeti dataset, the Rhino classifier performs poorly because of a very limited dataset compared to the other animals. Work on improving the Rhino classifier would be a big win.

  • tfimgsort needs to be generalized to accept any network. Right now it assumes InceptionV3. This is hardcoded in the tensor names it extracts for classification

Clone this wiki locally