augment_batch_iterator.py 文件源码

def get_scaled_translated_img_bb(self, name, bb):
        im = imread(name)
        img = cv2.cvtColor(im, cv2.COLOR_RGB2BGR)
        h0, w0, _ = self.img_size
        wc, hc = (bb[0] + bb[2]) / 2, (bb[1] + bb[3]) / 2
        face_width = (bb[3] - bb[1]) / 2
        # Old approach: scale and then translate
        res, new_face_width, shc, swc = self.compute_scale_factor(
            img, face_width, hc, wc)
        thc, twc = self.compute_translation(new_face_width, shc, swc)
        # New approach: translate and then scale
        # thc, twc = self.compute_translation(face_width, hc, wc,
        #                                     min_pad=self.MIN_FACE_SIZE + 10)
        # high_scale = np.min([thc - 5, h0 - thc - 5, twc - 5, w0 - twc - 5])
        # res, new_face_width, shc, swc = self.compute_scale_factor(
        #     img, face_width, hc, wc,
        #     high_scale=high_scale, low_scale=None)
        out_bgr, new_bb = self.copy_source_to_target(res, new_face_width,
                                                     shc, swc, thc, twc)

        log = "%.1f,%.1f,%.0f\n" % (
            (new_bb[1] + new_bb[3]) / 2, (new_bb[0] + new_bb[2]) / 2, new_face_width * 2)
        with open('aug.csv', mode='a', buffering=0) as f:
            f.write(log)
        # cv2.rectangle(out_bgr, (int(new_bb[0]), int(new_bb[1])), (int(new_bb[2]), int(new_bb[3])),
        #               (255, 255, 0), thickness=4)
        # cv2.imwrite("%d.jpg" % os.getpid(), out_bgr)
        # sys.exit(0)
        out = cv2.cvtColor(out_bgr, cv2.COLOR_BGR2RGB)
        return out, new_bb