Roboflow

Best Region Proposal Models

Compare the best 1 region proposal models and try them on your own image, free in the Roboflow Playground.

1 model · 1 open-weight · 1 free to try · prices synced Jul 28, 2026

Open-Source Region Proposal Models

1 model with downloadable weights you can self-host under their licenses (MIT). All run live in the Playground through hosted APIs, so self-hosting is optional.

Azure
Florence-2
Florence-2, introduced by Microsoft Research at CVPR 2024, is an open-source vision-language foundation model designed to unify diverse computer vision tasks within a single sequence-to-sequence framework. Unlike traditional models that specialize in specific tasks, Florence-2 accepts both images and text prompts and outputs text for tasks such as captioning, object detection, segmentation, OCR, and region-based grounding. It comes in two sizes—Florence-2-base (~230M parameters) and Florence-2-large (~770M parameters)—and is trained on FLD-5B, a large dataset of ~126M images with ~5.4B annotations.The model demonstrates strong zero-shot and fine-tuned performance, often rivaling larger vision-language systems while remaining lightweight and efficient. Released under the MIT license, all weights are publicly available, making it accessible for fine-tuning and deployment in applications like VQA, content tagging, accessibility, and research. Florence-2’s compact design, versatility, and openness position it as a practical alternative to larger proprietary multimodal models.

How to Choose a Region Proposal Approach

Region proposal generates candidate object regions scored by objectness rather than class. Today it is mostly a component inside detectors, so the real question is whether you need it standalone at all.

Two-stage detectors like Faster R-CNN already include a learned Region Proposal Network; you get proposals by using the detector, not by adding a separate system. Standalone proposal generation earns its place in weakly supervised labeling, candidate narrowing before an expensive model, and research settings where recall over all plausible objects matters more than classification.

If the goal is "find everything that might be an object", class-agnostic outputs from promptable segmenters and open-vocabulary detectors now cover much of what classical proposal methods offered, with better quality. Reach for classical proposals mainly when compute is tight or a pipeline explicitly expects them.

The bottom line: Use the proposal stage built into two-stage detectors, and for standalone candidate generation consider promptable or open-vocabulary models before classical methods.

Frequently Asked Questions About Region Proposal Models

Region proposal is the task of generating candidate boxes that are likely to contain some object, scored by objectness rather than class. It began with algorithms like selective search and became the learned Region Proposal Network (RPN) inside two-stage detectors such as Faster R-CNN, where proposals are then classified and refined. Proposal quality is measured by recall: whether true objects are covered by at least one candidate. Beyond detector internals, proposals are useful for weakly supervised labeling and for narrowing expensive downstream analysis to promising regions. This page lists 1 region proposal model, including 1 open-weight option you can self-host; all of them run live in the Playground so you can test them on your own images.

It depends on your task and constraints. For fixed categories in production, a model fine-tuned on your own data typically beats any general-purpose model. Compare the region proposal models on this page and try them on your own image to see which fits.

Yes. 1 of the 1 region proposal models here are open-weight (for example Florence-2), free to self-host under their licenses (MIT).

Yes. You can run all 1 in the Roboflow Playground for free. Upload an image and compare the models' output side by side, no setup required.

This page lists the one region proposal models in the Roboflow Playground catalog, all of them open-weight and free to self-host; all of them run live in the Roboflow Playground on your own images. Compare licenses, parameters, and release dates side by side, or open any model page for full details.