consnet.api.bbox

consnet.api.bbox.pair_iou(bboxes1, bboxes2)[source]

Compute the intersection-over-unions (IoUs) among human-object pairs.

Parameters:

  • bboxes1 (Tensor[N, 8]) – Human-object pairs to be computed. They are expected to be in (x1, y1, x2, y2, ...) format.

  • bboxes2 (Tensor[M, 8]) – Human-object pairs to be computed. They are expected to be in (x1, y1, x2, y2, ...) format.

Returns:

The computed pairwise IoU values

Return type:

Tensor[N, M]

consnet.api.bbox.pair_nms(bboxes, scores, method='fast', hard_thr=0.5, soft_thr=0.3, sigma=0.5, score_thr=1e-06)[source]

Perform non-maximum suppression (NMS) on human-object pairs. This method supports multiple NMS types including Fast NMS [1], Cluster NMS [2], Normal NMS [3] and Soft NMS [4] with linear or gaussian suppression terms.

Parameters:

  • bboxes (Tensor[N, 9]) – Batches of human-object pairs to be suppressed. The values are expected to be in (batch_id, x1, y1, x2, y2, ...) format.

  • scores (Tensor[N]) – Human-object interaction detection scores to be considered.

  • method (str, optional) – Type of NMS. Expected values include 'fast', 'cluster', 'normal', 'linear' and 'gaussian', indicating Fast NMS, Cluster NMS, Normal NMS and Soft NMS with linear or gaussian suppression terms.

  • hard_thr (float, optional) – Hard threshold of NMS. This attribute is applied to all NMS methods. Human-object pairs with IoUs higher than this value will be discarded.

  • soft_thr (float, optional) – Soft threshold of NMS. This attribute is only applied to linear and gaussian methods. Human-object pairs with IoUs lower than hard_thr but higher than this value will be suppressed in a soft manner.

  • sigma (float, optional) – Hyperparameter for gaussian method.

  • score_thr (float, optional) – Score threshold. This attribute is applied to normal, linear and gaussian methods. Human-object pairs with suppressed scores lower than this value will be discarded.

Returns:

Human-object pairs and their updated scores after NMS. The values are expected to be in (batch_id, x1, y1, x2, y2, ..., score) format.

Return type:

Tensor[N, 10]

References

  1. Bolya et al. (https://arxiv.org/abs/1904.02689)

  2. Zheng et al. (https://arxiv.org/abs/2005.03572)

  3. Neubeck er al. (https://doi.org/10.1109/icpr.2006.479)

  4. Bodla et al. (https://arxiv.org/abs/1704.04503)