This is a retrained version of the Faster R-CNN object detection network trained with the Common Objects in Context (COCO) training dataset. The actual implementation is based on Detectron, with additional network weight pruning applied to sparsify convolution layers (60% of network parameters are set to zeros).
The model input is a blob that consists of a single image of 1, 3, 800, 1280 in the BGR order. The pixel values are integers in the [0, 255] range.
| Metric | Value |
|---|---|
| Mean Average Precision (mAP) | 38.74%** |
| GFlops | 849.9109 |
| MParams | 52.79 |
| Source framework | TensorFlow* |
See Average Precision metric description at COCO: Common Objects in Context. The primary challenge metric is used. Tested on the COCO validation dataset.
Image, name: input, shape: 1, 3, 800, 1280 in the format B, C, H, W, where:
B- batch sizeC- number of channelsH- image heightW- image width
Expected color order is BGR.
The net outputs a blob with the shape 300, 7, where each row consists of [image_id, class_id, confidence, x0, y0, x1, y1] respectively:
image_id- image ID in the batchclass_id- predicted class ID in range [1, 80], mapping to class names provided in<omz_dir>/data/dataset_classes/coco_80cl_bkgr.txtfileconfidence- [0, 1] detection score; the higher the value, the more confident the detection is- (
x0,y0) - normalized coordinates of the top left bounding box corner, in the [0, 1] range - (
x1,y1) - normalized coordinates of the bottom right bounding box corner, in the [0, 1] range
[*] Other names and brands may be claimed as the property of others.
[**] May be different from the original implementation due to different input configurations.