Hi Henrik, we're doing a limited rollout after Monday's first release but I'm collecting names for users interested in signing up for an account. I can follow up at a later date with more information. Thanks!

That’s perfect, thank you Nicole

I hope so! There are some ethical questions that must first be addressed regarding releasing models that accurately find critically endangered or highly poached species, as well as legal agreements that would need to be worked out with each data partner regarding the use of their data to train public models. On the AI team we're pretty passionate about open sourcing as much as possible, so fingers crossed 🙂

Gotcha. Well I will stay tuned! It would certainly be a powerful combo.

If nothing else, we might be able to work out releasing models trained only on a subset of the data (albeit a large one) that doesn't include endangered species or data from privacy-concerned partners. Which species were blacklisted would need to be clearly communicated 🙂

I agree that the features in our networks have a huge data advantage over basically anything else out there besides possibly the megadetector, and that's class agnostic 🙂

Totally. That would be fantastic. Most of the Nature Conservancy's (org I work for) immediate classification needs/challenges don't involve endangered species, so a model trained on a subset of non-sensitive data would be perfect.

It won't be tomorrow, but I'll definitely try to work out how we can make that happen in the future.

You rock! Thanks @Sara Beery 🙏

Just spitballing, another option could be to export the graph without the final layers, so truly export a feature extractor as opposed to a classifier....

oh yes - that's what I was getting at! Apologies I was probably misusing ML terms... I'm new to the AI world

but yes I was curious about using the WI model's base and fine tuning it with our own labeled data. For orgs with a fair amount of labeled data (specific to the ecosystems they're working in and with classes that pertain to their business needs), I figure that might be a good approach for combating generalization challenges and stretching a relatively small quantity of training images.

Totally! If you're not doing it already, I'd recommend taking the Microsoft MegaDetector (https://github.com/microsoft/CameraTraps/blob/master/megadetector.md) running it over your images, and keeping the high-confidence boxes, then performing "distillation" by assuming your image class label matches the box label and fine-tuning the detector with your set of classes. You can do that right away, no waiting for lawyers required 🙂

GitHub

microsoft/CameraTraps

Tools for training and running detectors and classifiers for wildlife images collected from motion-triggered cameras. - microsoft/CameraTraps

Original URL: https://github.com/microsoft/CameraTraps/blob/master/megadetector.md

Or you can use a chained approach, where you assume the detections are correct and train a cropped-box classifier (like what @Mathias Tobler has been doing)

We described this high-level approach in a short paper at the KDD Data Mining and AI for Conservation Workshop this year: https://arxiv.org/abs/1907.06772

Sweet! This is all super helpful. Thanks for pointing me in the right direction Sara 👍

Hey @Sara Beery - the WI try-ai page seems to have disappeared. Do you know if that's moved elsewhere or if it's no longer available?

It was taken down due to some concerns about AI ethics, regarding classification of humans. WI is currently trying to determine how to handle these cases, but they decided to take down the model temporarily while they determine what, if any, action to take.

ah, gotcha. OK thanks for the update !

I don't know the answers, but I've posted these questions to the internal WI slack channel. I'll try to get answers for you ASAP (I also want to know 🙂 )

I do know that WI is currently only open to whitelisted users, but I believe you can join a waitlist to get whitelisted.

@Ben Weinstein Wildlife Insights is in a beta release right now. We’re approving accounts only for members of our Trusted Tester program which is at capacity for the moment. We do plan to reopen later in the year, so keep a lookout for an announcement! For those who aren’t in the Trusted Tester program, the account will not be recognized by the system and you won’t be able to reset your password. Unfortunately the messaging around this has been confusing - we’ll be working to improve the clarity. Sorry for the confusion!

Thanks for the clarification @Nicole Flores! What about how to best get involved?

I’d suggest signing up for our Trusted Tester program so we can reach out when we do open the program later this year https://www.wildlifeinsights.org/trusted-tester

Is there a system or process for people with machine learning or data science or modeling expertise to join the team and help out? As opposed to practitioners who might want to use the system.

I’ll have to check in with others who lead this work but will get back to you ASAP!

Funny, I just went through the "reset password" loop again this morning. Signed up for the Trusted Tester program months ago and never got any response, science committee seems to be dead too. Communication is definitely an issue here. At the same time people in and outside of our organization keep bringing up Wildlife Insights as a platform to manage our camera trap data. Hope things will open up soon.

Mathias, I know from the AI team side that everyone is working hard to get the platform to a point where it can be opened up to the community! The scale of this is massive, and that means it moves slowly. Thinking about communication (which isn't my job, but discourse is useful), would it be helpful to email an update and apology those who signed up for the beta testing program and are still waiting? Nobody likes being left hanging.

Sara, I understand that this is a huge effort and that it is better to get bugs worked out before an official launch. I think the PR department has gotten a bit ahead of the tech team in promoting the platform when it is not quite ready. I think a bit more communication would be helpful, especially for people who have been involved in this for a while but are not part of the core group.

I'll take this feedback back to the AI team and hopefully they can pass it to whoever can take action on some interim communication!

Hi Mathias – I really appreciate your input and I apologize that the beta launch and press has been misleading. We want to make sure the system can scale smoothly and that our team can provide support to all users when we do open to new users. Building the features needed and sorting out these processes so that miscommunication doesn’t happen in the future does take some time, and we thank everyone for their patience. I’m eagerly awaiting the next release of Wildlife Insights as much as everyone else! We’ll plan to send an update to the Trusted Tester sign ups and others involved when we have a firm roadmap for our next release.

Hi Nicole - thanks for the response. I am genuinely interested in the success of Wildlife Insights and I know that it is a huge effort. I have been involved in various planing meetings at several stages and and I really interested in contributing my experience with camera trap data management, databases, camera trap data modeling etc. as much as possible. However, sometimes it feels like it's hard to keep a foot in the door since we are not part of the core NGOs. Maybe that is part of my impatience 😉 Looking forward to the next update.