In the computer vision (CV) community, it is widely known that high-quality, diverse data is the key to achieving better model outcomes. So, you have high-quality data, which is leading to good models. The question now is: How do you take your good models to the next level? Consider the following four aspects of your data pipeline to transform your good models into great models:
How is your OPP?
We’re not talking about the 90’s rap song, but rather your Ontology, Protocols and Practices (OPP). If you don’t have your annotation ontology and accuracy standards clearly defined and documented, you’ll be forever drowning in wasted time, inconsistencies, and frustration with your annotated data. An OPP document sets the stage for your growing CV organization—defining what gets annotated and how. Get your team focused and aligned with an OPP today.
Quality Frame Indication and Processing
Video data is a compilation of scenes filled with objects, actions, and events over time. Unfortunately, the potential factors that could impact the visual quality (and therefore the usefulness of the frame) of the targets of interest to you in a scene are endless... poor lighting, large areas of dark shadows, camera movement, movement of objects, etc. If you are not considering a quality indicator at the frame, scene, or video level to evaluate its utility in training your models, then you could be impeding your ability to create great models.
Work Smarter with Pre-Annotation
Training models is a continually iterative process. Gathering high-quality data shouldn’t be. Are you using automation to speed up the process? Leveraging AI to pre-annotate your data can get you 85% of the way toward your dataset labeling needs. Taking the first pass of auto-labelling your data significantly reduces the time, effort, and money dedicated to getting high-quality data. Then your human team takes over to perform quality assurance and develop training datasets on the data where the automation needs to improve. This continuous improvement process will accelerate your efforts, reduce time, and lower your overall labor costs by at least 15-20%.
Data Diversity Quotient
Are you finding that your business use cases for CV are driving the creation of complex models (i.e., many classifiers, object detectors, deep feature extractors, etc)? If so, it is important to have a diverse collection of data in your training pipelines in order to ensure uniformity in performance. At Voxel51, we measure the quality of our datasets via a Data Diversity Quotient (DDQ), which goes beyond basic statistics like object/class counts and distributions and assesses the diversity of scene types, object appearances, and other higher-level semantic information about a dataset. Diverse datasets prevent models from overfitting during training data and, therefore, mitigates problems like reduced detection/classification accuracy due to domain transfer on your production data.
We’ll dive deeper into each of the above topics in future posts, so stay tuned!
High-quality, intentionally-curated data is everything when it comes to training great CV models. At Voxel51, we have over 25 years of CV experience and think deeply about achieving state-of-the-art outcomes for our customers. As a video-first company, we offer expertise, best practices, and a cutting-edge Platform that powers the full CV lifecycle, from image and video annotation to performing inference on videos at scale. Time is of the essence in bringing your AI-powered product to market. If you want to accelerate and strengthen your data pipelines to make it happen , let’s chat! You can reach us at email@example.com.