As someone who's annotated 1000+ images across detection, segmentation, and pose estimation tasks for multiple projects, Roboflow has become my default platform - not because of hype, but because it consistently removes friction at every stage of the CV pipeline.

The annotation tooling is genuinely best-in-class. The AI labeling suite covers Label Assist, Smart Polygon via SAM, Box Prompting, and Auto Label - each suited for a different stage of the project. For a new dataset with no existing model, Auto Label saves hours. For refinement, Smart Polygon with a single click gets you polygon masks in seconds. Rapid's new annotation control lets you go from AI-generated boxes to production-quality annotations without leaving the platform - that alone has cut my dataset preparation time significantly.

The export flexibility is underrated. Getting your dataset out in COCO, YOLO, Pascal VOC & many other formats with a single click, directly integrated into training scripts via the Python SDK, means zero pipeline glue code. The free tier is genuinely useful - not a crippled trial - and covers storage, annotation, and export for personal and research projects. They even provide you 3 credits which you can further use for fine-tuning your dataset directly on their GPUs, which saves your time if you don't have a good enough computation power locally.

Beyond the product: Roboflow's OSS contributions are real. RF-DETR and the YOLO weight releases aren't marketing - they're models I've actually benchmarked and deployed. Their community engagement on their socials is fast and technically substantive, not just support ticket deflection(shoutout to Trevor). Their engineers push improvements to their GitHub projects almost daily - shoutout to Piotr Skalski, the person who got me into CV in the first place. I’d also suggest you check out their blogs and demo videos; you can learn a lot from them.

If there's one area for improvement, it's that larger dataset operations (bulk re-export, version management at scale) can feel slow on the free tier. A minor friction point given everything else it offers. Review collected by and hosted on G2.com.

The main friction I've hit is around pricing transparency at scale.

The free tier is genuinely useful, but the transition to paid tiers involves a credit-based system where it's not always clear upfront how quickly credits deplete for operations like augmentation or Auto Label runs on large datasets. Users with large datasets also run into file size restrictions and slower performance, which I've noticed when working with high-resolution frames from video pipelines.

Additionally, advanced model training and deployment customization options are limited - if you want fine-grained control over training hyperparameters or custom deployment configurations, you'll quickly hit the ceiling and need to export to your own stack.

For a free-tier user doing research projects it's a minor issue, but teams building production pipelines should factor this in early. Review collected by and hosted on G2.com.

What I value most is how much Roboflow condenses into one platform. I can go from labeling a dataset to training and deploying without stitching together five different tools. The annotation workflow is genuinely fast, and the assisted labeling saves me a lot of time versus doing it manually. Best of all, I deploy my own custom models through Roboflow's serverless Hosted API, so I get a production endpoint without managing any infrastructure. It's also easy to use, so I'm not fighting the interface to get work done. Review collected by and hosted on G2.com.

If I had to nitpick, cold starts on the serverless Hosted API can add a bit of latency on the first request after idle time, so for latency-sensitive cases you have to plan around that. Pricing can also climb as inference volume grows, so it's worth keeping an eye on usage at scale. Neither has been a dealbreaker for how I use it, but they're the areas I'd like to see improve. Review collected by and hosted on G2.com.

Roboflow offers a strong solution, responsive personal support, and cutting-edge models for object detection, all while keeping the operating price reasonable. Review collected by and hosted on G2.com.

The basic tier doesn’t include a review platform, and the number of seats is limited. However, once you upgrade, they offer a great solution for external data taggers, which feels like a strong fit for growing companies. Review collected by and hosted on G2.com.

AI-powered road damage detection made efficient. Roboflow helps streamline our AI workflow for detecting road surface damage from vehicle-mounted images. The dataset annotation and augmentation tools improve model accuracy across different road conditions and lighting. We can quickly iterate and deploy models, which is essential for near real-time inspection. Overall, it significantly reduces development time and improves reliability in road condition analysis. Review collected by and hosted on G2.com.

While Roboflow is very effective for rapid development, advanced model customization and fine-tuning still require external tools. Additionally, pricing can scale up quickly when working with large datasets or high-volume usage, which may be a concern for long-term projects Review collected by and hosted on G2.com.

My favorite thing about Roboflow is how user-friendly it is (for the most part). I've never coded anything before, but needed to do some coding for a post-grad internship; specifically, I was tasked to create something that would automatically detect plastic in images. One of my contacts recommended Roboflow to me, and even though I have run into a couple of challenges, I've been able to do what I needed to do for my internship through Roboflow.

Additionally, I appreciate the support offered. The video tutorials helped me a lot when I was starting out, and, one I got further along, the AI assistant helped me polish my workflows.

Finally, I really appreciate one of the staff members, Gustavo Loureiro dos Reis. He has been so helpful in getting me to finish my project because he is very knowledgeable when it comes to the platform, responsive, and kind. Review collected by and hosted on G2.com.

The least helpful aspect of Roboflow is definitely how difficult it was to get support from an actual person. My company paid for the Core plan, yet we were not able to contact a real Roboflow staff member when we ran into an unsolvable issue. I reached out to the AI assistant, the community forum, and Roboflow support, but nothing worked. The AI assistant was unable to help, the community forum did not receive a response, and Roboflow support (represented by an employee named Lenny) informed us that we did not have access to support and then proceeded to ignore the emails I sent afterward.

The only reason we were able to get support from Gustavo is that after being subscribed to the Core plan for three months, we happened to receive an email offering us the opportunity to schedule a meeting and speak with a real person to review the platform.

Another thing I dislike about Roboflow is that it does not properly advertise how many credits training will take, which led to me exhausting all of mine very early in the month this month. I had been training all my models with custom weights, but decided to test out the new "Neural Architecture Search" feature recently because it was recommended. I had no idea it was going to train so many models and consume so many credits (which I believe is what happened) because that was not made clear beforehand. Review collected by and hosted on G2.com.

Building machine learning models on Roboflow is really easy. There are plenty of datasets available on Roboflow Universe, and you can even add your own images to improve your model—it’s super straightforward. the best part is that Roboflow handles the training for you and delivers great accuracy. Review collected by and hosted on G2.com.

Roboflow Univers doesn’t have as many datasets as Kaggle, but that makes sense since the platform is more focused on its own ecosystem. Training a model can take some time, depending on how large your dataset is—but the more data you have, the better the results. Overall, it’s a solid experience. Review collected by and hosted on G2.com.

What I like most about Roboflow is its end-to-end computer vision workflow, especially how smoothly it brings dataset management, annotation, preprocessing, augmentation, and export together in a single platform. Having access to a range of diverse, well-annotated public datasets (for example, 3D printing defect datasets), combined with flexible export formats and built-in version control, makes experimentation and reproducibility much easier. On top of that, Roboflow’s interface cuts down the overhead of dataset preparation, so researchers can spend more time on model design and evaluation instead of getting bogged down in data engineering. Review collected by and hosted on G2.com.

Roboflow is a highly comprehensive platform, and most of its features feel well thought out. If anything, some of the more advanced functionality takes time to fully explore and use effectively, but that’s largely because the platform includes such a broad range of powerful tools. Overall, that depth is a real strength: it lets users grow from straightforward projects into more advanced, research-oriented workflows as their needs evolve. Review collected by and hosted on G2.com.

What I find most valuable about Roboflow is how efficiently it allows me to annotate large datasets. Using their annotation tool, I was able to label thousands of images and complete an entire dataset in just three hours. This is a process that would have otherwise taken two to three days. I also appreciate how easy it is to organize projects and seamlessly export datasets directly into my Python notebook, which significantly streamlines my workflow Review collected by and hosted on G2.com.

Frankly, I do not currently dislike anything about Roboflow. if I do find something I dislike in the future, I will let you know. Review collected by and hosted on G2.com.

The best thing about Roboflow is that I can create my own computer vision models and, beyond that, configure them from scratch so they work exactly the way I need. The simple drag-and-drop features make it easy to set things up, and I can make intuitive changes on the go. Since it’s all online, it doesn’t depend on my PC’s cores. Review collected by and hosted on G2.com.

So far, it’s been great exploring Roboflow. Still, I’d like more utility when it comes to bringing in and managing our own data, especially for custom models that require a huge amount of data. If possible, it would also be helpful to have ways to procure or import data directly from different sites like Kaggle, etc. Review collected by and hosted on G2.com.

Roboflow provides an end-to-end pipeline for computer vision research, from dataset annotation and versioning to augmentation and model export. The interface is intuitive, and the dataset management features (splits, class balance visualization, augmentation control, and format conversion) significantly reduce experimental overhead. For academic research, the ability to quickly iterate on datasets, reproduce experiments, and export to multiple frameworks (YOLO, COCO, TensorFlow, ONNX, etc.) is extremely valuable. Review collected by and hosted on G2.com.

For very large-scale datasets, cloud processing and export can sometimes be time-consuming, and advanced customization of augmentation pipelines is slightly more constrained compared to fully custom local scripts. Additionally, some advanced features are locked behind paid tiers, which may be a limitation for students with limited funding. Review collected by and hosted on G2.com.