search

happyharrycn / actionformer_release

Code release for ActionFormer (ECCV 2022)
MIT License
430 stars 77 forks source link

why set num_classes as 1 in ActivityNet dataset module? #67

Closed fake-warrior8 closed 1 year ago

fake-warrior8 commented 1 year ago

Hi, I want to know why you set num_classes as 1 in ActivityNet dataset module line122, which make the gold label is a trivial zero. The setting comes from anet_tsp.yaml.

dataset: { json_file: ./data/anet_1.3/annotations/anet1.3_tsp_filtered.json, feat_folder: ./data/anet_1.3/tsp_features, fileprefix: v, file_ext: .npy, num_classes: 1, input_dim: 512, feat_stride: 16, num_frames: 16, default_fps: 15, trunc_thresh: 0.5, crop_ratio: [0.9, 1.0],

upsample the features to a fixed length of 192

max_seq_len: 192, force_upsampling: True, }

tzzcl commented 1 year ago

For ActivityNet v1.3, we learn action proposals rather than the real action instances due to the optimization issue. We comine the action proposals with external action classifiers to obtain the final action predictions, if you have more question, please refer to our paper .

fake-warrior8 commented 1 year ago

For ActivityNet v1.3, we learn action proposals rather than the real action instances due to the optimization issue. We comine the action proposals with external action classifiers to obtain the final action predictions, if you have more question, please refer to our paper .

I will check it. Thank you!