Meta released DETR under Apache 2.0 in 2020, and at 41.6 million parameters self-hosting is genuinely easy: PyTorch, the Hugging Face transformers library, and a GPU if you want it quick, though it runs without one. If you would rather not install anything, it runs in your browser on CNAPS Studio, included in the basic plan.
What made this model matter is how it does the job rather than how well. Detectors before it scanned an image in pieces, proposed thousands of candidate boxes, and then threw most of them away with hand-written cleanup rules. DETR replaced all of that with 100 slots that each go looking for one object, filled in a single pass, with no candidate boxes and no cleanup step. Six years on, that design is why the newer detectors in this catalog have DETR in their names.
The cost of being first is that it has been overtaken. It scores 42.0 average precision on the standard detection benchmark, and a newer model in the same catalog reports 54.7 on the same measure. Three detectors and two specialised fine-tunes sit in the CNAPS Studio catalog, and this is the one to reach for when you want the general-purpose original rather than the fastest option.
What it actually is
Four parts in a line. A convolutional backbone reads the image into features, a transformer encoder builds a view of the whole picture at once, a transformer decoder runs 100 object queries against that view, and two small heads turn each query into a class label and a box. The 100 queries are a hard ceiling: it cannot return more than 100 objects from one image.
Training used COCO 2017, a public set of 118,000 annotated photographs covering 91 everyday categories, for 300 epochs. The matching step during training is the clever part, pairing each prediction to at most one real object, which removes the need for duplicate cleanup afterwards.
Output is a plain list. For every object found: a class label such as person, dog, car or chair, a box given as left, top, right and bottom coordinates, and a confidence score from 0 to 1. That list is the whole product, which makes the model a first step rather than a finished answer.
Parameters41.6 millionArchitectureEncoder-decoder transformer over a convolutional backboneObject queries100, which is also the maximum objects returned per imageClasses91 everyday categories from COCO 2017Training dataCOCO 2017, 118,000 annotated photographsTraining run300 epochs, three days on sixteen data-centre GPUs, batch size 64OutputClass label, box as left/top/right/bottom, confidence 0 to 1PublishedMay 2020LicenseApache 2.0Downloads last month613,233
The published numbers
The model card reports 42.0 average precision on the COCO 2017 validation set. That metric is scored out of 100 and rewards both finding the right objects and drawing tight boxes around them, averaged over how strictly the box has to match. Training took three days on sixteen data-centre GPUs at a total batch size of 64.
That 42.0 comes from the authors, so it is vendor-reported. The counterweight is that COCO average precision is the most re-run number in computer vision and DETR is six years old, so the figure has been reproduced independently many times, a far stronger position than a recent model quoting itself. Adoption agrees: 613,233 downloads in the last month, more than any other detector here.
One number in our own documentation does not match: our page says 150 training epochs where the model card says 300. The model card is the source of record, and the figure above follows it.
How it compares to the other detectors here
DETRRF-DETR MediumYOLO-STypeTransformer set-prediction detectorReal-time transformer detectorLightweight real-time detectorSize41.6M, runs on a CPU, faster on a GPU33.7M, runs on a small inference GPULightweight; see its own postClasses91 everyday COCO categories80 everyday COCO categories91 everyday COCO categoriesCOCO average precision42.0, author-reported54.7, measured in-house by RoboflowSee its own postOutputLabelled boxes with confidence scoresLabelled boxes with confidence scoresLabelled boxes with confidence scoresLicenseApache 2.0Apache 2.0See its own postDistinctiveThe original set-prediction design, and the most downloaded detector here4.4 ms per image at 576 by 576The lightest option in the catalog
This is the rare comparison where the numbers really are on the same scale, and there is still a catch. DETR's 42.0 and RF-DETR Medium's 54.7 are both COCO average precision on the same validation split, but the second was measured in-house by Roboflow rather than taken from each model's own paper, so treat the gap as the right direction and roughly the right size rather than an exact margin.
The practical choice is about what you are optimising. Reach for DETR when you want the well-understood general detector and speed is not the constraint, for RF-DETR Medium when you need real-time throughput, since it is both more accurate and far faster, and for YOLO-S when you want the lightest option over the same categories. When the thing you are hunting is not on the 91-category list at all, SAM 3.1 (Scene) finds it from a text phrase instead, though it returns masks rather than labels.
What to chain it with
It takes one image and returns a list of labelled boxes with confidence scores. Almost everything useful about it comes from what reads that list next.
Three chains cover most of it. Image Crop by Class cuts out every object of one label, so a shelf photo becomes one image per bottle, which then goes to Object Classification for a finer verdict. Image Masker by Class turns the detections into a mask instead, feeding Selective Blur for anonymising a frame or PowerPaint for removing what you found. Image Gate by Text passes an image onward only when a given label is present, turning the detector into a batch filter rather than an annotator.
Open DETR in CNAPS Studio, run a photo with more than a hundred things in it, and watch where the 100-query ceiling bites before you build anything on top of it.
Sources
- huggingface.co/facebook/detr-resnet-50 for the Apache 2.0 license, the 41.6 million parameter count, the 42.0 average precision on COCO 2017 validation, the 300 training epochs, the training hardware and batch size, the 100 object queries, the dataset size, and the monthly download count.
- github.com/facebookresearch/detr and arXiv:2005.12872 for the architecture and the set-prediction training method.
- github.com/roboflow/rf-detr for RF-DETR Medium's 54.7 average precision and the note that Roboflow measured every row in-house.
- docs.cnaps.ai/ai-model-hub for the CNAPS Studio input and output and the other detectors in the catalog.