RM新时代有限公司

Recently, CloudWalk Technology's CongRong Large Model has made another important breakthrough in the visual field. The object detector based on the visual foundational large model stood out from many well-known enterprises and research institutions such as Microsoft Research (MSR), Shanghai Artificial Intelligence Laboratory, and Beijing Academy of Artificial Intelligence on the famous benchmark COCO dataset in the detection field, setting a new world record.

CloudWalk Technology's CongRong Large Model achieved a mean Average Precision (mAP) of 0.662 on the COCO test set, ranking first in the list (see figure below). On the validation set, it achieved a single-scale mAP of 0.656 and a multi-scale TTA mAP of 0.662, both reaching world-leading levels.

Big Data Combined with Self-Supervised Learning to Build Core Visual Technology

Big data self-supervised pre-training, represented by GPT, has achieved remarkable breakthroughs in the field of natural language understanding (NLP). In the visual field, significant progress has also been made in training foundational models with big data combined with self-supervised learning.

On one hand, extensive visual data helps the model learn general foundational features. CloudWalk's visual foundational large model uses over 2 billion data points, including a large number of unlabeled datasets and image-text multimodal datasets. The richness and diversity of the datasets enable the model to extract robust features, greatly reducing the development cost of downstream tasks.

On the other hand, self-supervised learning does not require manual annotation, making it possible to train visual models on massive amounts of unlabeled data. CloudWalk has made various improvements to self-supervised learning algorithms, making them more suitable for fine-grained tasks such as detection and segmentation. Achieving good results on the COCO detection task is proof of this.

Open Object Detection + Zero-Shot Detection Capability Significantly Reduces R&D Costs

Thanks to the excellent performance of the visual foundational model, CloudWalk's CongRong Large Model can support zero-shot detection of thousands of object categories based on large-scale image-text multimodal data training, covering various objects in industries such as energy, transportation, and manufacturing.

Zero-shot can mimic the human reasoning process, using past knowledge to infer the specific form of new objects in the computer, thereby enabling the computer to recognize new things.

How to understand zero-shot? Suppose we know the morphological characteristics of donkeys and horses, and that tigers and hyenas are striped animals, while pandas and penguins are black-and-white animals. We define a zebra as a horse-like animal with black and white stripes. Without looking at any pictures of zebras, just by reasoning, we can identify zebras among the animals in a zoo.

CloudWalk's visual foundation large model demonstrates strong generalization performance, greatly reducing the data dependency and development cost required for downstream tasks. At the same time, zero-shot significantly improves training and development efficiency, making widespread application and rapid deployment possible.