A picture is worth a thousand words.
Trex-compressed.mp4
- T-Rex is an interactive object counting model that can first detect then count any objects through visual prompting, which is highlighted by the following features:
- Open-Set: T-Rex possess the capacity to count any object, without constraints on predefined categories.
- Visual Promptable: Users can provide visual examples to specify the objects for counting.
- Intuitive Visual Feedback: T-Rex is a detection-based model that allows for intuitive visual feedback (i.e. detected boxes), enabling users to assess the accuracy of the result.
- Interactive: Users can actively participate in the counting process to rectify errors.
🔥 We release the training and inference code and demo link of DINOv, which can handle in-context visual prompts for open-set and referring detection & segmentation. Check it out!
- T-Rex provides three major workflows for interactive object counting / detection.
- Positive-only Prompt Mode: T-Rex can detect then count similar objects in an image with just a single click or box drawing. Additional visual prompts can also be added for densely packed or small objects
- Positive with Negative Prompt Mode: To address false detections caused by similar objects, users can correct the detection results by adding negative prompts to the falsely-detected objects.
- Cross Image Prompt Mode: This feature supports counting across different reference and target images, ideal for automatic annotation. Users only need to prompt on one reference image, and T-Rex will detect objects in other target images. Note that this feature is still under development, and the performance is not guaranteed.
- T-Rex can be applyed to various domains for detection/counting including but not limited to Agriculture, Industry, Livestock, Biology, Medicine, Retail, Electronic, Transportation, Logistics, Human, etc.
- T-Rex can also serve as an open-set object detector, which can be applied for automatic annotation. It process exponential zero-shot detection capability, and offers strong performance in dense and overlapping scenes.
- We list some of the potential applications of T-Rex below:
@misc{jiang2023trex,
title={T-Rex: Counting by Visual Prompting},
author={Qing Jiang and Feng Li and Tianhe Ren and Shilong Liu and Zhaoyang Zeng and Kent Yu and Lei Zhang},
year={2023},
eprint={2311.13596},
archivePrefix={arXiv},
primaryClass={cs.CV}
}