Sailing towards Amazon of AI: Step 2

Name of Project

Publish ultra low latency Object Detection Network for face detection and tools on the Ocean Market enabling privacy tech related business models.

Description

Evotegra GmbH provided the Object Detection Network and high performance inference engine that was used during the AI premiere in German live television show “Bundespolizei Live” provided by Kabel1 on 18/09/2019. During the 2 hour show our network processed 50 Full-HD television pictures per second on a single consumer grade GPU with low latency to reliably detect and obscure all faces in the image. This enabled us to meet the particularly high demands on the protection of individual privacy in cooperation with the Federal Police.

Typically the TensorRT conversion of a model introduces a regression in the model performance. We adapted the tools and process to assure a lossless TensorRT conversion of the frozen graph and support both full floating point precision (FP32) and half floating point precision (FP16).

What is the problem solved?

With our proposal we want to evaluate the market response to not just datasets but high quality AI algorithms and tools.

What is the final product/result

We propose to create and publish a distributable package that contains:

  1. Pretrained Tensorflow 1 Object Detection Model optimized for ultra low latency. The model can be retrained to different purposes.
  2. Tools for lossless TensorRT conversion
  3. Documentation of the conversion process
  4. A restricted inference engine to validate the capabilities of the model

The package will be distributed as binaries only within a downloadable Docker image.
The system requirements are:

  • X86_64 (AMD64) plattform
  • NVIDIA GPU (Pascal or higher)
  • Docker

The package does not contain:

  • an unrestricted inference engine
  • support beyond the tools and samples

How does this project drive value to the Ocean ecosystem?

As a first mover this new type of data related asset will extend the scope of the Ocean Market as a whole. Instead of being just a data market it will contribute to transform Ocean Protocol into the “Amazon of AI” and as such into a one stop shop for data, algorithms and turn key solutions.

Sailing towards the Amazon of AI our proposal would strengthen the Ocean Marketplace as valuable resource for AI/ML professionals by expanding the Ocean ecosystem from datasets to algorithms. Nevertheless we see this just as an intermediate step. In the upcoming DAO round we then hope to present the obvious next step.

ROI

Our intial target price is 1500 USD. We strive to mint 20 tokens. Based on a 10/90 pool the initial liquidity required for an AMM pool is:

1500 * 20 / 9 = 3333 USD ~ 4000 Ocean (as of 28/02/2021)

With a 10/90 pool about 3333 USD would flow directly back into TVL for the creation of a new liquidity pool which is pledged not to be removed for a minimum period of 6 month and not more than 5 pools shares per week will be removed after that period.

Acting as a reference point for future publications we think it will attract 10x - 100x more similar data related assets and such has a potential to return 33,000 USD (40,000) to 333,000 USD (400,000 Ocean) in TVL.

Licence

This proposal will be published under a commercial use license with restriction to redistribution

Important notes

We implemented the provided network for Face Detection in mere 4 weeks. While it was obviously fit for use in Live-TV this network has significant potential for improvement. As a provider of customer specific perception solutions we are there to help with the adoption to any arbitrary purpose. Also customers interested in licensing our high performance inference engine can contact us via the bottom contact link.

One part of the grant goes into creating a downloadable distributable package.
The initial target price for this algorithm in the market is EUR 1500. We strive to mint 20 datatokens. As we pledge to not withdraw our liquidity from the pool for a minimum 6 month period, it would not be possible for us to provide the necessary liquidity within a 70/30 pool. Yet we would provide the initial liquidity to create a 10/90 pool with the help of the ocean community, as we are not familiar yet with Ocean.py or similar scripts.

Proposal Wallet Address

0x61B15998893cC746B46C08FEdEE13a0d1b33bBa9

Short Bio

Tobias Manthey
Role: CEO & founder Evotegra, Regional Manager German AI Association, Teamlead AI, Software Developer
Twitter: @evotegra
LinkedIn: https://www.linkedin.com/in/tobias-manthey-3b232928/

Background/Experience:
Founder at Evotegra GmbH
The initial focus of Evotegra GmbH in 2001 was software integration in the asset management and investment banking industry. In 2013 I published my traffic sign detection app AcoDriver which shifted our focus to AI and image processing in automotive and industry. Today we are a provider for individual visual perception solutions based on AI/Deep Learning with a focus on embedded systems and real-time processing.

Articles and Links:

See our network and engine in action in the YouTube recording at:
30:07, 30:51, 31:34, 36:11, 36:33, 1:06:00


Our press release with YouTube video markers.

How to contact us:

2 Likes