Learning by Aligning Videos in Time
We present a self-supervised approach for learning video representations using temporal video alignment as a pretext task, while exploiting both frame-level and video-level information. We leverage a novel combination of temporal alignment loss and temporal regularization terms, which can be used as supervision signals for training an encoder network. Specifically, the temporal alignment loss (i.e., Soft-DTW) aims for the minimum cost for temporally aligning videos in the embedding space. However, optimizing solely for this term leads to trivial solutions, particularly, one where all frames get mapped to a small cluster in the embedding space. To overcome this problem, we propose a temporal regularization term (i.e., Contrastive-IDM) which encourages different frames to be mapped to different points in the embedding space. Extensive evaluations on various tasks, including action phase classification, action phase progression, and fine-grained frame retrieval, on three datasets, namely Pouring, Penn Action, and IKEA ASM, show superior performance of our approach over state-of-the-art methods for self-supervised representation learning from videos. In addition, our method provides significant performance gain where labeled data is lacking.
Our Solution
Our RetroActivityTM platform automatically builds computational models of a complex physical task, such as an assembly activity, given only a handful of recorded demonstrations of the task.

Once such a model is built, RetroActivity finely tracks the job status from live video, to guide an operator through the task, provide independent training, and perform analytics.

RetroActivity offers audible and visual alerts to help the operator avoid assembly mistakes whereas its analytics capability identifies non-value added activities across processes, unexpected variability in process times, and traces assembly mistakes through a line.

Benefits
RetroActivity boosts quality control and continuous improvement by digitizing manual assembly activities through visual analytics, providing increased visibility to industrial engineers, line operators, operational excellence leaders, and trainers.
Continuous Improvement
Quality
Control
Operator
Training
Product
Ramp Up

Differentiation
Poka Yoke Systems
Physical poka-yoke mechanisms such as limit switches or color-coding help workers avoid errors in "real-time", yet they have limited applicability and their coverage becomes even shallower when multiple configurations are coming down the same line. RetroActivity is a versatile solution that intervenes to prevent a much larger variety of potential mistakes, and doesn't require re-designing a process to apply.
Machine Vision for Visual Defect Detection
Visual part inspection systems focus on Quality Control stations, which are generally found at the end of the assembly line. Our customers find that by then it is too late: workers have already made their mistakes, and the product either needs to be disassembled and fixed or be discarded, which are both costly options. In fact, some steps in correct assembly cannot be captured by looking at the final product at all and thus it becomes important to check at the source if the worker missed a step or performed a step incorrectly. In contrast, RetroActivity enables continuous quality checks while workers are performing manual assembly.
Video Analytics
Some vendors sell offline human activity understanding to provide "descriptive" analytics of workstation processes. We differentiate ourselves from them in three ways:

1. Real-time Guidance versus Offline Analytics: We provide live task guidance and immediate worker feedback, directly adding value by conveniently mistake-proofing customers' processes as opposed to competitors' after-the-fact analytics solution.

2. Rapid Onboarding: We give a visual querying tool to industrial engineers allowing them to build QA experiences within 4-6 hours as opposed to 4-6 weeks that competitors take.

3. Prescriptive analytics versus Descriptive Analytics: Competing solutions "describe" what industrial engineers have already observed on the line. In contrast, our solution understands the 3D scene and canonical human action to retrieve, refine, and recommend processes improvement ideas i.e. action items for industrial engineers.

Digital Work Instructions
DWI is one of several platforms providing "interactive PowerPoint" capabilities that allow industrial engineers to rapidly convert paper-based work instructions to digital work instructions. It may sometimes be possible to integrate machine vision based part defect detection into such solutions as well.

1. Throughput Reduction: DWI requires operators to place a part being assembled under a camera, then press a button on the screen to scan the part, and get permission from the system to move to the next step, and do so at every step of the process. This slows down the line.

2. Change in operator behavior: The hardest part of digitization in manufacturing is changing human behavior. We find that operators get annoyed when they must meet a daily quota and are being held back by the machine; and make attempts at bypassing the process altogether. In contrast, our solution proactively observes the worker and product and interrupts only when necessary.

3. Redesign of Assembly process: DWI requires a complete redesign of the assembly process with DWI software at its center needing manual data entry or even QR codes on products. Our solution adapts naturally to existing processes.
Install without stopping your line
We drop by to help you install, or you can do it yourself within an hour.
Onboard within hours
Capture
Enter a bill of process and record around 25 demonstrations of the process (takes ~1 hour)
Label
Label the videos at the level of individual steps (takes ~10 minutes)
Deploy
Provide live feedback to assembly operator or passively analyze video for improvements.
Analyze
Get cycle times and step-level analytics, such as standard deviation for each step

Schedule a demo
Get access to the leading video analytics platform for your factory floor.
Phone: +1 669-220-8352
E-mail: info@retrocausal.ai

Retrocausal, Inc.
8201 NE 164th Ave, Suite 200
Redmond, WA 98052
Made on
Tilda