DEtection TRansformer (DETR) model trained end-to-end on COCO 2017 panoptic (118k annotated images). It was introduced in the paper End-to-End Object Detection with Transformers by Carion et al. and first released in this repository.

Disclaimer: The team releasing DETR did not write a model card for this model so this model card has been written by the Hugging Face team.

The DETR model is an encoder-decoder transformer with a convolutional backbone. Two heads are added on top of the decoder outputs in order to perform object detection: a linear layer for the class labels and a MLP (multi-layer perceptron) for the bounding boxes. The model uses so-called object queries to detect objects in an image. Each object query looks for a particular object in the image. For COCO, the number of object queries is set to 100.

The model is trained using a "bipartite matching loss": one compares the predicted classes + bounding boxes of each of the N = 100 object queries to the ground truth annotations, padded up to the same length N (so if an image only contains 4 objects, 96 annotations will just have a "no object" as class and "no bounding box" as bounding box). The Hungarian matching algorithm is used to create an optimal one-to-one mapping between each of the N queries and each of the N annotations. Next, standard cross-entropy (for the classes) and a linear combination of the L1 and generalized IoU loss (for the bounding boxes) are used to optimize the parameters of the model.

DETR can be naturally extended to perform panoptic segmentation, by adding a mask head on top of the decoder outputs.

You can use the raw model for panoptic segmentation. See the model hub to look for all available DETR models.

Here is how to use this model:

Currently, both the feature extractor and model support PyTorch.

The DETR model was trained on COCO 2017 panoptic, a dataset consisting of 118k/5k annotated images for training/validation respectively.

The exact details of preprocessing of images during training/validation can be found here.

Images are resized/rescaled such that the shortest side is at least 800 pixels and the largest side at most 1333 pixels, and normalized across the RGB channels with the ImageNet mean (0.485, 0.456, 0.406) and standard deviation (0.229, 0.224, 0.225).

The model was trained for 300 epochs on 16 V100 GPUs. This takes 3 days, with 4 images per GPU (hence a total batch size of 64).

This model achieves the following results on COCO 2017 validation: a box AP (average precision) of 40.2, a segmentation AP (average precision) of 31.9 and a PQ (panoptic quality) of 44.6.

For more details regarding evaluation results, we refer to table 5 of the original paper.