retinanet-tf#
Use Case and High-Level Description#
RetinaNet is the dense object detection model with ResNet50 backbone, originally trained on Keras*, then converted to TensorFlow* protobuf format. For details, see paper, repository.
Steps to Reproduce Keras* to TensorFlow* Conversion#
Clone and install the original repository (tested on
47fdf189
commit) following the installation steps. Make sure you ranpython setup.py build_ext --inplace
from the root of repository.Download the original model from here
Get conversion script:
Get conversion script from repository:
git clone https://github.com/amir-abdi/keras_to_tensorflow.git
Checkout the commit that the conversion was tested on:
git checkout c841508a88faa5aa1ffc7a4947c3809ea4ec1228
Apply
keras_to_tensorflow.patch
from<omz_dir>/models/public/retinanet-tf
:
git apply keras_to_tensorflow.patch
Run script:
python keras_to_tensorflow.py --input_model=<model_in>.h5 --output_model=<model_out>.pb
NOTE: The conversion was tested with keras==2.3.1, tensorflow==1.15 and h5py<3.0.0.
Specification#
Metric |
Value |
---|---|
Type |
Object detection |
GFlops |
238.9469 |
MParams |
64.9706 |
Source framework |
TensorFlow* |
Accuracy#
Metric |
Value |
---|---|
coco_precision |
33.15% |
Input#
Original Model#
Image, name: input_1
, shape: 1, 1333, 1333, 3
, format: B, H, W, C
, where:
B
- batch sizeH
- image heightW
- image widthC
- number of channels
Expected color order: BGR
.
Mean values: [103.939, 116.779, 123.68]
Converted Model#
Image, name: input_1
, shape: 1, 1333, 1333, 3
, format: B, H, W, C
, where:
B
- batch sizeH
- image heightW
- image widthC
- number of channels
Expected color order: BGR
.
Output#
Original Model#
Classifier, name:
filtered_detections/map/TensorArrayStack_2/TensorArrayGatherV3
. Contains predicted bounding boxes classes in a range [1, 80]. The model was trained on Common Objects in Context (COCO) dataset version with 80 categories of objects, 0 class is for background. Mapping to class names provided in<omz_dir>/data/dataset_classes/coco_80cl_bkgr.txt
fileProbability, name:
filtered_detections/map/TensorArrayStack_1/TensorArrayGatherV3
. Contains probability of detected bounding boxes.Detection box, name:
filtered_detections/map/TensorArrayStack/TensorArrayGatherV3
. Contains detection boxes coordinates in a format[y_min, x_min, y_max, x_max]
, where (x_min
,y_min
) are coordinates of the top left corner, (x_max
,y_max
) are coordinates of the right bottom corner. Coordinates are rescaled to input image size.
Converted Model#
The array of summary detection information, name - DetectionOutput
, shape - 1, 1, 300, 7
in the format 1, 1, N, 7
, where N
is the number of detected
bounding boxes. For each detection, the description has the format:
[image_id
, label
, conf
, x_min
, y_min
, x_max
, y_max
], where:
image_id
- ID of the image in the batchlabel
- predicted class ID in range [1, 80], mapping to class names provided in<omz_dir>/data/dataset_classes/coco_80cl_bkgr.txt
fileconf
- confidence for the predicted class(
x_min
,y_min
) - coordinates of the top left bounding box corner (coordinates stored in normalized format, in range [0, 1])(
x_max
,y_max
) - coordinates of the bottom right bounding box corner (coordinates stored in normalized format, in range [0, 1])
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.txt
.