README Document

The codes are based on mmdetection. If you have never used the mmdetection framework before, please read the orignal README document.

Template Detector

The detector takes two images as inputs: template image and fact image. You can see the detector code for details. Load config file: template_detector_r50.py

Other implementation of the detector is similar to other mmdetection detectors.

• Overview The detector is based on RetinaNet, which is a multi-class RPN, i.e., single-stage detector. If you are not familiar with the RetinaNet, please read the paper for details.

• Backbone The backbone is ResNet-50. You can change the backbone to ResNet-101 in the config file. The input images of the backbone are template and fact images. It outputs the features of both images.

• Neck The neck part of RetinaNet in mmdetection is FPN. Original FPN take feature list of one image as the input. Here modified TemplateFPN takes two feature lists. You can write your codes with the merge_feats method to hanlde the feature merge process.

• Heads The two heads, which are namely subnets in the RetinaNet paper, are still used in the detector for classification and regresssion.

Data

Write your dataset file like COCO and CustomDataset classes in TemplateDataset file.

Annotations

[
    {
        'template_img': 'a.jpg',
        'fact_img': 'b.jpg',
        'width': 1280,
        'height': 720,
        'bboxes': <list> [[x1, y1, w, h], ...]
        'labels': <list> [1, ...]
    },
    ...
]