search

HYPJUDY / Decouple-SSAD

Decoupling Localization and Classification in Single Shot Temporal Action Detection
https://arxiv.org/abs/1904.07442
MIT License
95 stars 19 forks source link

About mAR #16

Open faith30 opened 5 years ago

faith30 commented 5 years ago

Hi, yupan.I got a strange result mAR:0.0. The eval code is https://github.com/pratik18v/sst/blob/5e4d3f1e34824223a5f7a28cd5c40a8f0fba420e/code/recall-eval/recall-eval.ipynb which is a common script for mAR.I used nr_propsosals=1000,tio=(0.5,0.95,0.05).The proposal file is result_fuse_UCF101_decouple_ssad.txt which you give and the annotation file is Decouple-SSAD/data/thumos14/test/thumos14_test_annotation.csv which you also give.

HYPJUDY commented 5 years ago

Hi, could you explain why do you evaluate the proposals with mAR/recall? I'm not familiar with mAR metric since we use mAP metric in action detection.

faith30 commented 5 years ago

Hi! It's for retrieving more accurate proposals in an acceptable number of generating proposals. The metric more emphasis the importance of the proposal generation methods,such as bsn and tag...