Launch: Superior Dataset Search Filters, Operators, and Logic

Dataset high quality and the flexibility to simply curate your information are essential to constructing an efficient laptop imaginative and prescient mannequin. The better it’s so that you can search and discover your information, the higher you possibly can curate your dataset to enhance mannequin efficiency.

We’re excited to announce superior dataset search filters, operators, and logic, accessible now in all Roboflow workspaces. These options allow you to raised discover and perceive your dataset in any respect levels of mannequin constructing, from getting ready your first dataset model to creating incremental enhancements as your mannequin is utilized in manufacturing.

The brand new filtering options accompany the prevailing semantic search capabilities in Roboflow dataset search. These capabilities mean you can seek for an summary key phrase (i.e. “transport container”) and discover associated photos. Now, you possibly can each question utilizing a semantic search and slim your filter with our superior filters.

On this information, we’re going to present you the right way to use the brand new dataset search options accessible within the Roboflow utility to curate datasets for constructing laptop imaginative and prescient fashions. With out additional ado, let’s get began!

Introducing Superior Dataset Search

Contemplate a situation the place your mannequin struggles to determine one class despite the fact that you might have labeled many photos with that class. With an efficient dataset exploration instrument, you possibly can study your current photos to reply questions like “are my photos too comparable?” and “are there unannotated situations of this class?”

Roboflow’s new search characteristic makes answering such questions – and plenty of different questions you might have about your dataset – simpler than ever. Within the Roboflow utility, now you can search photos by:

• Filename
• Tags
• Picture width and top
• The variety of annotations in a picture
• The lessons current on the picture (or exclude photos with a labeled class)
• The cut up a picture is in

You may mix search options utilizing AND or OR statements, permitting you to construct advanced queries to discover your dataset.

Listed here are a few of the many questions you possibly can reply with the brand new dataset search options:

1. What number of photos include an object that isn’t labeled?
2. What number of photos include a particular mixture of lessons?
3. What number of photos exist that characteristic a selected class in your legitimate check set?
4. What number of photos include fewer than two annotations?
5. What number of photos include a particular label and have no less than three annotations?

Let’s stroll by the right way to use the dataset search after which present just a few examples. To see a full reference checklist of search capabilities in Roboflow, confer with the Search a Dataset documentation.

To entry the brand new dataset search, click on on the Photographs tab within the sidebar of a venture in your workspace. Then, click on the search bar above the pictures on the web page. This search bar is enabled with our new dataset search options.

While you open the search bar, a number of instance “operators” will seem. An operator is an attribute by which you’ll be able to question.

Let’s run just a few queries. For this information, we’ll use the Microsoft COCO dataset, which accommodates over 120,000 photos. First, suppose we wish to discover all photos that include a cat and a canine, two lessons in our dataset. We are able to discover them utilizing the next question:

class:cat AND class:canine

Above, there are lots of examples of photos with labeled cats. We may make a extra particular question and filter by cut up (i.e. solely present photos with annotated cats and canine within the coaching check set), filename, and the opposite attributes talked about above.

Let’s run one other check. Suppose our mannequin performs poorly at figuring out cats. We are able to run a question to search for all photos that don’t include a “cat” annotation however do include a cat. We are able to achieve this by leveraging the semantic search capabilities constructed into the Roboflow search characteristic. 

While you specify a key phrase for which to look (i.e. “cat”), Roboflow will order search outcomes based on their relevance to that key phrase. We do that utilizing vector embeddings. We calculate an embedding on your textual content question (i.e. “cat”) and evaluate it to the picture embeddings on your dataset. We then return the outcomes whose embeddings are closest to the question.

The next question will allow us to discover photos the place we now have missed annotating a cat:

-class:cat cat

This question excludes all photos that include a “cat” class then searches for photos related to the textual content question “cat”. Listed here are the outcomes:

We are able to click on by to a picture to discover every picture:

On this picture, there’s a “eating desk” label however no label for the cat. An annotation was missed through the labeling course of. We may repair the annotation and repeat this course of for various lessons to scrub up the dataset.

Suppose we wish to search for photos which have the category “cellular phone” and no less than three annotations on the picture in complete. We may achieve this utilizing the next question:

class:"cellular phone" min-annotations:3

The search question efficiently returned photos that characteristic class “cellular phone” and include no less than three annotations. Be aware: The min-annotations search flag counts all annotations, not annotations in a particular class.

Search Filters

Under are the search filters accessible on the time of launching this characteristic. Confer with the Search a Dataset documentation for the most recent updates on superior dataset search in Roboflow.

• like-image:<SOURCE_ID>: Type by semantic similarity measured by CLIP.
• tag: Filter by user-provided tags.
• filename: Runs a seek for file names that match the supplied file title. Use * initially and finish of a question to run a partial match.
• cut up: Filters by cut up (practice, check, legitimate).
• job:<JOB_ID>: Exhibits photos with the supplied job ID.
• min-width:X: Exhibits photos with a width lower than X.
• max-width:X: Exhibits photos with a width higher than X.
• min-height:X: Exhibits photos with a top lower than X.
• max-height:X: Exhibits photos with a top higher than X.
• min-annotations:X: Filters photos with greater than the desired variety of annotations.
• max-annotations:X: Exhibits photos with fewer than the desired variety of annotations.
• class:CLASS: Exhibits photos which have no less than one annotation with the supplied label.
• -class:CLASS: Exhibits photos that don’t characteristic a selected class

You may mix these attributes above utilizing AND or OR statements.

Conclusion

The brand new Roboflow search characteristic, accessible to be used now, gives a strong suite of options to be used in looking out and exploring a dataset. With the options described above, you’ll find photos that meet a standards, determine photos which might be lacking labels, and discover photos that characteristic a selected piece of metadata (i.e. a filename that accommodates a string, or a picture with a tag).

This characteristic is on the market to all customers on free and paid plans. In case you are all in favour of storing giant, personal datasets in Roboflow to be used with our superior dataset search characteristic, contact the Roboflow gross sales staff to be taught extra about pricing.