How to calculate the AP value as shown in paper

[4-0-4-notfound]

Hello, i want to reproduce your result in the paper. The evaluation tool needs r , t, and score to get the final AP result. I want to know how to get these values from your code. It seems the cluster_mat_pred cluster_center_pred is the rand t, but how to get the score?

Could you provide the example code?

[lvwj19]

Thanks for your attention. The score is the predicted visibility.

[4-0-4-notfound]

1、for calculating the AP, is it need to filter the vs_picked_idx = pred_vis_val > 0.45?

2、And what is pred_conf_val used for?

3、which kind of projectionType is used for generetate GT, orthogonalor perspective?

4、it seem the t is need to scale 1000 to get AP? self.dataset['data'] *= scale (1000)

[lvwj19]

1. Yes, it is needed and the threshold can be adjusted by yourself. pred_vs_val is used to filter the ungraspable instance, i.e., score.
2. pred_conf_val is used to filter the point-wise prediction before voting and you can refer to our journal paper.
3. We used perspective.
4. You can check it and just keep the predictions and labels consistent on the scale.

[4-0-4-notfound]

Thanks for your response. So pred_conf_val is not involved in calculating AP?

[lvwj19]

Yes, you can ignore it if you haven't trained the conf branch, because it is the post-processing optimization.

[4-0-4-notfound]

Thanks a lot

[lvwj19]

There are the codes that you can refer to.

   # filter objects with vs lower than vs_threshold
    vs_threshold = 0.45
    pred_results = list(zip(pred_vs_cluster, cluster_center_pred, cluster_mat_pred))
    pred_results = [ rst for rst in pred_results if rst[0] > vs_threshold ]
    pred_results.sort(key=lambda x:x[0], reverse=True)

    result_list = []
    for rst in pred_results:
        tmp_dict = {}
        tmp_dict['score'] = 1.0 * rst[0] 
        tmp_dict['t'] = (rst[1] / 1000.0).tolist()
        tmp_dict['R'] = rst[2].tolist()
        result_list.append(tmp_dict)