Looking for resources on object detection

Part 1 (2017)
1

I just finished up with lesson 7 after going back-and-forth through the material several times to make sure I really understood what is going on, and doing the fisheries competition got me really interested in object detection given that there are multiple fish in many of the images.

Instead of simple classification, I’d like to be able to paint a box around EACH fish in an image, and even be able to identify different kinds of fish in a single image.

I’m using Python3, Keras 2,and TF … looking for any recommendations on tutorials, resources, best practices, etc… that might be helpful. I’d love to be able to apply whatever is out there to both images and video.

1 Like
2

I’m starting to do the same thing. The simplest path for me is to just use Tensorflow’s object detection, it’s very thorough and they release pre-trained weights. They have a tutorial on how to fine-tune an existing network for a new dataset: https://github.com/tensorflow/models/blob/master/object_detection/g3doc/running_pets.md

There’s also this very fun tutorial: https://medium.com/towards-data-science/how-to-train-your-own-object-detector-with-tensorflows-object-detector-api-bec72ecfe1d9

I’m curious what other people use for the actual image annotation. Many seem to be python/GTK-based desktop apps. But I’d love a tool that could be run as a webserver with a decent mobile UI so I could annotate images myself when I have some free time, but not need to be in front of the computer. Any recommendations?

3

I am also looking at this. Couple of links that I added to my list for further study:

4

Thanks for the links.

Are you running the TF object detection locally or are you doing it in the cloud?

5

Have you tried the SSD port?

I’ve seen that link before and it looks promising although I can’t say definitively how it compares with R-CNN and YOLO implementations (which I also see frequently mentioned).

6

Locally, I built a DL box with a GTX 1070 GPU for this kind of stuff.

7

@simoneva I found a much better Keras port of SSD:

It provides comprehensive documentation and all of the original trained models.

8

Thanks. You might want to check out mask-rcnn which only came out recently. Quite complex but results on image segmentation are amazing.

9

Yeah, that paper is on my reading list. Thanks!

10

@wgpubs here is a repo on fish detection using yolov5

1 Like