hbornfree

Thanks for this! I have the exact use-case and have been using a Python script to do this for a while.

Hello HN,

This is Harsha, from the team at 3difyMe. We rolled out the first version of 3difyMe a week ago and wanted to get some feedback from HN community.

<TLDR>

3difyMe converts flat photos of T-shirts and other apparel into 3D models. You can generate multiple combinations by changing the {human, pose, background, footwear} and click unlimited photos or videos. These are ready to go to your online store or marketing platforms.

</TLDR>

—— Longer version aka Why are we doing this?

Circa Nov 2021, we were toying around with realistic 3D human models and wondered what business problems we could solve. Avatars were one (thanks to the incoming metaverse). But we didn't know if the timing was right. So we explored further and stumbled upon the E-commerce photography space. At the moment, apparel retailers typically use photos that are: 1. Flat-lay 2. Shot on a mannequin 3. Shot by professional agencies

Cost, effort, time and quality increase as you move down this list. A typical agency will charge you $15/product with a minimum order of 20 products and deliver in 2 weeks.

At 3difyMe we are trying to go from 1 to 3 using tech. No minimum order. And 1 minute to get results.

Here are our hypotheses about this: Set 1: > Smaller retailers doing flat-lay will be able to improve their visuals significantly with little effort. > Bigger retailers should be able to reduce spend, time and effort for a slight trade-off on quality.

Set 2: The ‘per image’ limit is irrelevant in a virtual studio. You’re free to take as many photos as you want. If retailers had a lot of photos, they could: > A/B test their visuals > Improve their marketing game with additional content at no extra cost

Set 3: In 2022, consumer attention is fully engaged in video (TikTok, YouTube, Netflix) and retailers barely have access to this medium. That’s because apparel video shoots are difficult and expensive. But not in 3D! > The next innovation cycle in retail is video and 3difyMe could make it accessible.

—— A note about the open source we’ve used: 3difyMe is built with Blender, ThreeJS and React. It’s just about amazing and unbelievable that these masterpieces are free and maintained so actively. We plan to contribute back both in code and funding in the near future. We ♥ you Blender and ThreeJS.

—

Feedback of all kinds is welcome. On product, market and product-market fit. :)

If you want to personally reach out, please write to us at harsha@3dify.me and team@3dify.me.

What's happening behind the scenes? Without knowing that it's hard to take the claim of 'no server involvement' at face value.

Does the sender have to be up for the receiever to receive?

Given how impossible it is achieve global consensus on policy, I'm pretty inclined to believe that a technological breakthrough is the only solution. Maybe a series of them.

In the best case, carbon removal and sequestration tech. In the neutral case, clean and efficient energy tech. In the worst case, climate-catastrophe coping tech.

This does resonate with me!

Happy to discuss and help you where possible. How can I reach out to you? Or maybe you can send me a 'hi' at harsha.xg{at}gmail.com?

Thanks for the suggestions @warrenm.

While I understand the importance of climate awareness and community building, I'm afraid that isn't where I shine.

Anything more technical, maybe in the fields of camera-traps, drone deployment, remote-sensing, satellite imagery, GIS, data analysis/viz etc may be a better fit. Trying to look for a way to get started in these.

Thanks again for taking the time to reply. Really appreciate it!

We are interested to solve the reverse problem of finding the mask of text given an image with text on it. Any ideas? Machine-learnt or otherwise.

+1 Will help many devs who do remote work and have to take calls.

We haven't decided on the pricing nor implemented it in code. That's the next thing to do. Probably we'll roll out both bundles ($x for y images) or subscriptions.

As for the copy, you're right. We need a lot of improvement on targeting. We thought LSPs and marketers are going to be main customers. Have to refine there. Our first idea was to do an API but we deferred it as the output still needs a pair of human eyes to approve.

Thank you for the feedback and wishes. Really appreciate it!

> takes many hours to manually do this right now

This was also the original problem we set out to solve. Thank you for letting us know that more people share the same problem. Unfortunately we haven't yet gotten it into the hands of the marketing agencies.

If you are ok with it, can you send us a 'hello' mail at team{at}imgtranslate.com so that we can discuss a little more with you.

Cheers

It does OCR ofcourse. Pipeline is roughly:

OCR -> Text property identification -> Inpainting -> Region averaging -> Translation -> Fitting

Please have the LSP try it out and see if offers some value.

By integrating the tool to professional CAT tools we're guessing. Businesses almost always never buy machine translated output

Thank you for the feedback! Will fix it right away

It isn't. ANd as mentioned in the blog post, we want to discuss how it works in the next post and open source as much as we can. We're built on open-source stuff after all. :)

Yes we do know of Project Naptha. Infact I personally used it a couple of years ago. imgtranslate also has a text region identification module (through clustering mechanisms) but I have to admit it's not that great at the moment. So I've hidden it.

As far as inpainting goes, we should soon be switching to GANs which produce structurally more accurate and less blurry results.

The problem with photoshop is that - You need a lof of fonts installed. - It doesn't always render the non-latin fonts correctly - It's expensive

What you're describing sounds like the Google Translate camera feature(erstwhile WordLens). If you haven't tried it out, you should. Works quite well on photos taken on phone for signboards etc.

We are little particular about what to translate and how to faithfully reproduce it. That's the main differentiation.

Hello everyone! The creator of imgtranslate here. Happy to discuss anything on tech and product. The current stack uses a good dose of traditional(OpenCV) and neural implementations. And we're actively pursuing LSPs(Language Service Providers) and marketing agencies to try it out.

This is terrific! Very coincidentally, we were thinking of implementing a sentence de-noiser using sequence-to-sequence models only today evening. I work in the NLP domain writing Machine Translation systems. But NLP parsers are accurate for grammatically correct sentences only which necessitates the need for something like deep text correcter. Thank you for this. Will try this out this week and let you know how it goes.

Not Caffe. We're using a light weight in-house library which compiles and runs easily on mobile devices too. The advantage is primarily in size as we don't need as many features as a full blown framework like Caffe.

A friend did say that about the quote. I wanted to get more opinions. Will have to take it down today then. Thanks for the feeddback though. :)

The main programmer here. For the technically curious, the ingredients of the image processing are OpenCV and CNNs(Convolutional neural nets).

On design aspects, we wrote a short post here: http://blog.taglev.com/design-philosophy-behind-taglev/

Yes, even we found Caffe to be overwhelming. This library seems promising considering there are no dependencies and is quite portable to Android or iOS. https://github.com/nyanp/tiny-cnn

As others pointed out, Tesseract with OpenCV (for identifying and cropping the text region) is quite effective. On top of that, Tesseract is fully trainable with custom fonts.

In our use case, we've mostly had to deal with handwritten text and that's where none of them really did well. Your next best bet would be to use HoG(Histogram of oriented gradients) along with SVMs. OpenCV has really good implementations of both.

Even then, we've had to write extra heuristics to disambiguate between 2 and z and s and 5 etc. That was too much work and a lot of if-else. We're currently putting in our efforts on CNNs(Convolutional Neural Networks). As a start, you can look at Torch or Caffe.