This model is for a pedestrian detector used for Retail scenarios. The model is based on a backbone with hyper-feature + R-FCN.
| Metric | Value |
|---|---|
| AP | 80.91% |
| Pose coverage | Standing upright, parallel to image plane |
| Support of occluded pedestrians | YES |
| Occlusion coverage | <50% |
| Min pedestrian height | 80 pixels (on 1080p) |
| Max objects to detect | 200 |
| GFlops | 12.4 |
| MParams | 3.2 |
| Source framework | Caffe |
Average Precision (AP) is defined as an area under the precision/recall curve. Validation dataset consists of about 50,000 images from about 100 scenes.
-
name:
data, shape: [1x3x544x992] - An input image in the format [1xCxHxW]. The expected channel order is BGR. -
name:
im_info, shape: [1x6] - Image information [544, 992, 992/frame_width, 544/frame_height, 992/frame_width, 544/frame_height]
- name:
cls_prob_reshape, shape: [200, 2] - Softmax output across two classes [Background, Pedestrian] - name:
proposal: shape: [200, 5] - Object proposals [0, x1, y1, x2, y2] - name:
bbox_pred_reshape, shape: [200, 8] - Bounding box deltas [d0, d1, d2, d3] for two classes [Background, Pedestrian]. The refined bounding box is [(cx + w * d0) - 0.5 * w * exp(d2), (cy + h * d1) - 0.5 * h * exp(d3), (cx + w * d0) + 0.5 * w * exp(d2), (cy + h * d1) + 0.5 * h * exp(d3)], where w = (x2 - x1 + 1), h = (y2 - y1 + 1), cx = (x1 + 0.5 * w), cy = (y1 + 0.5 * h).
[*] Other names and brands may be claimed as the property of others.