Regarding the current development method of deep learning AI-based object detection models, the initial training cost is very high because it requires a large amount of image data for training depending on the detection target, and the cost of re-training is also high due to the decrease in performance (detection rate/recognition rate) caused by subtle environmental differences. Depending on the application (e.g., factories, cattle sheds, and pigsties), it takes a lot of time to acquire image data for training because it is difficult to access the site and often there are many cases where the image data itself cannot be exported.
With the current development method of the deep learning AI-based object detector, workflow including data acquisition, refinement, training, evaluation, deployment, and monitoring is not automated, making development time consuming. While some cloud-based workflow operations are automated, the automation solutions that can be applied on site are still insufficient. Furthermore, the development is repetitive and counterproductive as the deep learning model must be redeveloped from the start and a number of optimization techniques must be applied if there is a change in the environment or detection target.
SNUAILAB intends to provide an object detection solution capable of on-site adaption (self-training). This solution consists of a detector (DX) capable of detecting changes in target objects (e.g., livestock, crops, equipment/robots) in various domains (e.g., farms, factories), a training server (TX) to support detector performance monitoring and re-training the model if necessary, and remote assistant (see the figure below). Processing speed optimization is very important for the object detector (DX) as it performs primary inference on object class/location before performing associated recognition function (action, situation, face/plate-number, etc.) Therefore, SNUAILAB wishes to research and develop the company’s own model of a new structure that detects objects without anchor and NMS, unlike the existing anchor-based YOLO/SSD method, aiming to improve the processing speed by 50% and with an accuracy that is equivalent or higher compared to SOTA.
Swipe to right or left
SNUAILAB intends to provide a site-oriented AutoML platform for deep learning model development and application. SNUAILAB plans to provide a Vision AI-based AutoCare platform which minimizes repetitive and unproductive manual work required for model development and supports workflow automation that enables online learning and semi-supervised learning with images acquired on site. Especially, with the AutoCare platform, SNUAILAB aims to support rapid development of a lightweight custom detection model in the form of Edge AI, required in harsh environments.
Swipe to right or left
Swipe to right or left
Development of an object detection model with a new structure (Object Detection without Anchor)
Applied model compression algorithm which effectively reduces operation cost (model lightweighting through channel pruning)
Post-processing technique that requires a smaller amount of computation compared to NMS-based post-processing (Object Detection without NMS)
Swipe to right or left
Active flow for on-site self-training (Training Flow + Auto Label Flow)
Active learning, supervised learning, semi-supervised learning model applied for labeled/unlabeled data.
Swipe to right or left
Online Learning : model 𝑓𝑡⟶ data 𝐱𝑡⟶ estimate 𝑓𝑡(𝐱𝑡) ⟶ observation 𝑦𝑡 ⟶ loss 𝑙(𝑦𝑡, 𝑓𝑡(𝐱𝑡))⟶ model 𝑓𝑡+1
Semi-supervised Learning : labeled data and unlabeled (pseudo-labeled) data considered together in training