po0ya
commented
6 years ago Hey, feat_stride: 4 requires a higher resolution feature map than conv4_3, else we will be predicting two different values from the same location. Since conv4_3 feature map dimensions are 1/8 of the input's.
Maybe a new module on conv3_3. I think we tried it and it didn't make that much of a difference. It's worthwhile to do an analysis to see what is "recall vs size" of the smallest scale detection module for gt boxes with dim less than 8 with and without the new module.
On Wed, Jun 6, 2018, 7:50 AM loackerc notifications@github.com wrote:
Dear @mahyarnajibi https://github.com/mahyarnajibi and @po0ya https://github.com/po0ya, I was wondering if during implementation, have you guys experimented adding another detection layer at conv3_3 which has feat_stride: 4? Since wider face consists a lot of small faces. If yes, could you please share some insights? many many thanks.
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/mahyarnajibi/SSH/issues/28, or mute the thread https://github.com/notifications/unsubscribe-auth/AA9CpMRusPPCzOYcTL9rCsvp-9aknGRbks5t58HrgaJpZM4UcgJD .
Dear @mahyarnajibi and @po0ya, I was wondering if during implementation, have you guys experimented adding another detection layer at conv3_3 which has feat_stride: 4? Since wider face consists a lot of small faces. If yes, could you please share some insights? many many thanks.