Voxel51

What do you like best about the product?

It keeps the data clean and properly segregated, and it helps with labeling while improving the model’s performance on test benchmarks.

What do you dislike about the product?

It’s slightly slow at times when I’m running it locally.

What problems is the product solving and how is that benefiting you?

It’s helping me build better ML models, reduce turnaround time, and increase my overall productivity.

What do you like best about the product?

The MCAT format focuses on data annotation, and it’s geared toward physical AI.

What do you dislike about the product?

The interface could be more modern and more intuitive.

What problems is the product solving and how is that benefiting you?

It can support pre-training and experimenting with different ideas, models, and datasets.

What do you like best about the product?

I like that Voxel51 brings the training environment and sample datasets together in a single console. I like its user interface and value the sample datasets. Setup was quite easy and very intuitive.

What do you dislike about the product?

Limited data sources at the moment but I expect that to expand

What problems is the product solving and how is that benefiting you?

I use Voxel51 for training datasets, and it brings the training environment and sample datasets together in a single console.

What do you like best about the product?

What I appreciate most is 51’s ability to link visual data with sample-level metadata, which allows me to move beyond aggregate metrics and dive into individual cases with ease. In my robotics perception workflow, I use it to gain a unified view of RGB/RGB-D observations, segmentation masks, and grasp candidates.

Its integration with Python fits perfectly into my existing workflow; I can maintain my established perception and logging methods while simultaneously preparing outputs for visual inspection. This is particularly valuable when troubleshooting—for instance, if a grasp candidate looks questionable, I can quickly determine whether the issue stems from the segmentation results, the geometry, or incomplete observational data.

What do you dislike about the product?

In my robotics workflow, the primary challenge lies in customized integration. Conducting post-hoc analysis requires precise data mapping across RGB-D observations, overlaid grasp candidate regions, and event metadata. Furthermore, I am currently exploring ways to effectively correlate this information with synchronously collected tactile data and robot state data.

It would be highly beneficial to have more examples—specifically within the robotics domain—that demonstrate how to organize multimodal observational data and link frame-level annotations to failure events during task execution. Such resources would be invaluable to researchers, particularly those working with data that extends beyond the scope of standard image datasets.

What problems is the product solving and how is that benefiting you?

My robot perception workflow generates a variety of outputs, including standalone images, segmentation masks, depth data, grasp candidates, and event logs in JSON format. Correlating these outputs becomes a challenge when troubleshooting anomalous results.

To facilitate review within the FiftyOne platform, I have developed a structured export process that links visual outputs with sample metadata. This allows me to prepare cases for review in a consistent manner, enabling a direct comparison between raw observational data and the outputs of the perception pipeline.

Currently, the tangible benefit of this approach is a reusable integration mechanism for reviews, eliminating the need for manual preparation on a case-by-case basis. As this workflow is still in the validation phase, I cannot yet quantify its specific impact in terms of time savings, cost reduction, or improvements to actual grasp success rates.

What do you like best about the product?

It's beneficial in our workflow to validating datasets gaps/issues

What do you dislike about the product?

Good-sized learning curve on how to use it most efficiently

What problems is the product solving and how is that benefiting you?

It's helping us explore our datasets and improve our library of data

What do you like best about the product?

AI agent curation and model benchmarking as well as OOD data visualization via UMAP is cool

What do you dislike about the product?

Too long of a talk about Corso’s talk about his recent work.

What problems is the product solving and how is that benefiting you?

Saving money by preventing bad data from being annotated or to focus on data that’s more useful for the model

What do you like best about the product?

What I like most is that FiftyOne isn’t just a dataset viewer, it’s a real framework you can build on. For my Voxel51 × Twelve Labs hackathon project (CoachMe, an AI sports-coaching plugin), I built eight custom operators directly using FiftyOne’s plugin system. That operator framework let me connect video embeddings, similarity search, and AI feedback inside the App without having to build a UI from scratch.

I also liked being able to store per-sample fields—like embeddings, similarity scores, and coaching validation results as native dataset fields, and then explore them visually in the App. That made the whole data-curation workflow feel tight and cohesive. Between the visual explorer and similarity indexing, I could quickly tag near-duplicates and spot coverage gaps in a reference video library.

For a small team moving fast, the plugin architecture and the App’s out-of-the-box visualization saved me from writing a ton of infrastructure.

What do you dislike about the product?

The biggest friction for me was the learning curve around the plugin/operator system. The core dataset and App concepts are easy to pick up, but building custom operators required digging through the docs and existing plugin examples just to understand the execution context, the input/output schemas, and how to correctly write results back to samples. Under hackathon time pressure, this was the slowest part: I spent a lot of time trial-and-erroring operator behavior that clearer error messages or better examples would have shortened. Debugging was also tricky when an operator failed silently or when a field didn’t persist the way I expected, because it wasn’t always obvious what went wrong. For someone building their first FiftyOne plugin, more end-to-end operator examples and tighter error feedback would make the ramp-up much smoother.

What problems is the product solving and how is that benefiting you?

FiftyOne solved the data-curation and model-evaluation problem for my project specifically, turning a messy pile of video into a structured, queryable, visual dataset. For CoachMe, the core issue was that a coaching reference library built from random clips is “garbage in, garbage out”: duplicate camera angles, missing techniques, and no clear way to spot coverage gaps. With FiftyOne, I could store embeddings, similarity scores, and quality metadata as native fields, then build a similarity index to tag near-duplicates and surface gaps visually. That let me end up with five well-curated reference videos instead of fifty random ones better results with less data.

The bigger benefit was iteration speed. Rather than building dataset management, a visualization UI, and an evaluation layer from scratch, I got all three from FiftyOne and could focus on the actual product: the coaching operators. For a two-person team working against a hackathon deadline, that made the difference between shipping a working plugin and not shipping at all. The plugin framework also meant the work lived inside a tool other FiftyOne users could install and run, instead of being stuck as a one-off script.

What do you like best about the product?

I like FiftyOne for its ease of use and the accuracy of its agentic AI. The digital governance it offers is great, especially with its 'think different' change agents. It's straightforward to personalize and create math models with it.

What do you dislike about the product?

not applicable

What problems is the product solving and how is that benefiting you?

I find that FiftyOne solves finite math problems at Voxel. It offers ease of use, accuracy of agentic AI, and supports digital governance with innovative change agents.

What do you like best about the product?

I use FiftyOne for embedding, geotagging, and evaluations for agriculture, and it streamlines the process while helping me maintain logging. I like the UI and its open-source nature, which makes it easy to add extensions. I've built custom tooling for my specific workflow that helps me connect and retrieve data daily and send emails to my team. The initial setup was good mostly.

What do you dislike about the product?

I find the API and new feature compatibility for adding VLMs could be improved.

What problems is the product solving and how is that benefiting you?

I use FiftyOne for embedding, geotagging, and evaluations for agriculture. It streamlines my process and helps maintain logging. I love the UI and open-source nature, which lets me build custom tooling for workflow, connecting and retrieving data daily, and sending emails to my team.

What do you like best about the product?

I love using FiftyOne as the central orchestration layer for our computer vision pipeline. It's a total game changer for running evaluations on a model's predictions and instantly visualizing false positives and negatives in a high fidelity UI. The one-stop-shop functionality allows me to perform deep dive inspections of our ground truth annotations and verify model performance visually. It helps in fabricating high-quality models by ensuring the training data is clean, diverse, and representative of the actual engineering environments we monitor. The initial technical setup was remarkably efficient, and it effectively eliminates the friction of switching between platforms, helping me stay focused on creating quality models.

What do you dislike about the product?

While the core features are top tier, I find that the UI for the platform's more advanced features can feel like a bit of a departure from the rest of the software. But there is a noticeable incline in difficulty when you need to modify and set up custom features for a large and sophisticated project. Navigating the deeper configuration menus can sometimes feel like a journey in itself.

What problems is the product solving and how is that benefiting you?

FiftyOne solves the massive fragmentation of the AI development lifecycle and reduces context switching between disconnected tools, improving productivity.