ikuinen / CMIN_moment_retrieval

Cross-Modal Interaction Networks for Query-Based Moment Retrieval in Videos
86 stars 20 forks source link

Recall@5 Results #11

Closed AmeenAli closed 4 years ago

AmeenAli commented 4 years ago

Hello, I have trained a model and during the training I can only see the results for Recall@1 Mteic, can't see the results for Recall@5 metric , anyidea why?

ikuinen commented 4 years ago

This could be a reference with NMS. "backup_box" is all the candidate moments.

top_n_box = np.empty([top_n, 2])
top_n_box[:, 0] = backup_box[:, 0]

for i in range(1, top_n):
    num_backup = backup_box.shape[0]
    gt_starts1 = np.repeat(
        np.expand_dims(top_n_box[i - 1, 0], 0), num_backup, 0).reshape([-1])  # [backup]
    gt_ends1 = np.repeat(
        np.expand_dims(top_n_box[ i - 1, 1], 0), num_backup, 0).reshape([-1])  # [backup]
    backup_start = backup_box[:, 0].reshape([-1])
    backup_end = backup_box[:, 1].reshape([-1])
    IoUs = criteria.calculate_IoU_batch((gt_starts1, gt_ends1), (backup_start, backup_end))
    IoUs = IoUs.reshape([bsz, num_backup])
    keep_index = IoUs < tau
    backup_box = backup_box[keep_index]
    top_n_box[:, i] = backup_box[:, 0]
SCZwangxiao commented 3 years ago

Could you tell us what's your NMS threshold, i.e, the tau in the above code?