请问如何使用i3d特征

[twinsyssy1018]

你好，我看代码的数据处理，把特征都对齐为num_prop*400，然鹅i3d是2048维的特征，不知道应该如何处理。麻烦提供下使用i3d的pipeline ，谢谢

[wangxiang1230]

num_prop

Thanks for your attention to SSTAP, I3D features are in the format of "Length*Channel", where 100 is the temporal length and 2048 is the channel dimension. Similarly, for TSN features, 100 is also the temporal length and 400 is the channel dimension.

[twinsyssy1018]

谢谢，我重新训练下模型，但是在数据处理时出现iou的问题呢,我的start，end time 归一化后是0.675 0.85625,在计算 iou_with_anchors时（代码dataset.py的100行：tmp_gt_iou_map = [iou_with_anchors）] ） 出现匹配不上，最后gt_iou_map的size为0，报错了

[wangxiang1230]

谢谢，我重新训练下模型，但是在数据处理时出现iou的问题呢,我的start，end time 归一化后是0.675 0.85625,在计算 iou_with_anchors时（代码dataset.py的100行：tmp_gt_iou_map = [iou_with_anchors）] ） 出现匹配不上，最后gt_iou_map的size为0，报错了

Sorry, maybe I don't quite understand your question? It should be noted that the difference between TSN and I3D features is only in dimension (400 vs. 2048), you only need to change "feat_dim" (opt.py) from 400 to 2048, and change the path of dataset.

[twinsyssy1018]

谢谢回复，我的确没有改其他的地方，这个问题我再仔细看看代码吧~~