Skip to content

Latest commit

 

History

History
110 lines (70 loc) · 4.46 KB

README.md

File metadata and controls

110 lines (70 loc) · 4.46 KB
Raw

detr-resnet50

Use Case and High-Level Description

The detr-resnet50 model is one from DEtection TRansformer (DETR) models family, which consider object detection as a direct set prediction problem. The model has ResNet50 backbone and pretrained on Common Objects in Context (COCO) dataset for solving object detection task. DETR predicts all objects at once, and is trained end-to-end with a set loss function which performs bipartite matching between predicted and ground-truth objects. DETR simplifies the detection pipeline by dropping multiple hand-designed components that encode prior knowledge, like spatial anchors or non-maximal suppression.

More details provided in the paper and repository.

Specification

Metric Value
Type Object detection
GFLOPs 174.4708
MParams 41.3293
Source framework PyTorch*

Accuracy

Accuracy metrics obtained on Common Objects in Context (COCO) validation dataset for converted model. Background label + label map with 80 public available object categories are used (original indexing to 91 categories is preserved).

Metric Value
coco_orig_precision 39.27%
coco_precision 42.36%

Input

Original model

Image, name - input, shape - 1, 3, 800, 1137, format B, C, H, W, where:

  • B - batch size
  • C - number of channels
  • H - image height
  • W - image width

Expected color order is RGB.

Mean values - [123.675, 116.28, 103.53]. Scale values - [58.395, 57.12, 57.375].

Converted model

Image, name - input, shape - 1, 3, 800, 1137, format B, C, H, W, where:

  • B - batch size
  • C - number of channels
  • H - image height
  • W - image width

Expected color order is BGR.

Output

Original model

  1. Bounding boxes, name: boxes, shape - 1, 100, 4. Presented in format B, A, 4, where:

    • B - batch size
    • A - number of detected anchors

    For each detection, the description has the format: [x, y, w, h], where:

    • (x, y) - coordinates of the bounding box center(coordinates are in normalized format, in range [0, 1])
    • w - width of bounding box(values are in normalized format, in range [0, 1])
    • h - height of bounding box(values are in normalized format, in range [0, 1])

  2. Scores, name: scores, shape - 1, 100, 92. Contains scores in logits format for 91 Common Objects in Context (COCO) object classes. The last class is no-object class.

Converted model

  1. Bounding boxes, name: boxes, shape - 1, 100, 4. Presented in format B, A, 4, where:

    • B - batch size
    • A - number of detected anchors

    For each detection, the description has the format: [x, y, w, h], where:

    • (x, y) - coordinates of the bounding box center(coordinates are in normalized format, in range [0, 1])
    • w - width of bounding box(values are in normalized format, in range [0, 1])
    • h - height of bounding box(values are in normalized format, in range [0, 1])

  2. Scores, name: scores, shape - 1, 100, 92. Contains scores in logits format for 91 Common Objects in Context (COCO) object classes. The last class is no-object class.

Download a Model and Convert it into OpenVINO™ IR Format

You can download models and if necessary convert them into OpenVINO™ IR format using the Model Downloader and other automation tools as shown in the examples below.

An example of using the Model Downloader:

omz_downloader --name <model_name>

An example of using the Model Converter:

omz_converter --name <model_name>

Demo usage

The model can be used in the following demos provided by the Open Model Zoo to show its capabilities:

Legal Information

The original model is distributed under the Apache License, Version 2.0. A copy of the license is provided in <omz_dir>/models/public/licenses/APACHE-2.0-FacebookResearch.txt.