This is a person reidentification model for a general scenario. It uses a whole body image as an input and outputs an embedding vector to match a pair of images by the Cosine distance. The model is based on RMNet backbone that was developed for fast inference. A single reidentification head from the 1/16 scale feature map outputs the embedding vector of 256 floats. The model is provided without last calibration layer, but can be used in the same way as the original model (with insignificant drop in accuracy).
| Metric | Value |
|---|---|
| Pairwise accuracy | 92.93% |
| Pose coverage | Standing upright, parallel to image plane |
| Support of occluded pedestrians | YES |
| Occlusion coverage | <50% |
| GFlops | 0.12 |
| MParams | 0.82 |
| Source framework | Caffe* |
Pairwise accuracy is defined as the accuracy of the binary classification problem, where a positive class means that a pair of images represents the same person. Validation dataset consists of 10,000 image pairs of about 1500 persons.
-
name: "data" , shape: [1x3x160x64] - An input image in the format [BxCxHxW], where:
- B - batch size
- C - number of channels
- H - image height
- W - image width
The expected color order is BGR.
- The net outputs a blob with shape: [1, 256, 1, 1] named descriptor, which can be compared with other descriptors using the Cosine distance.
[*] Other names and brands may be claimed as the property of others.