Hi there, i am re-implementing dual shot using keras. During testing, my positive boxes scores are high, but the negative boxes also score quite high, NMS in this case cannot filter boxes so well, can someone enlighten me why this happens?
During training, i calculate the loss using all positive anchors with iou>0.4, negative anchors with iou<0.3. negative anchors are selected using hard negative mining, they are 3 times the number of positive anchors.
During testing, i use all initialized anchors to do a forward pass and do NMS on all anchors
Hi there, i am re-implementing dual shot using keras. During testing, my positive boxes scores are high, but the negative boxes also score quite high, NMS in this case cannot filter boxes so well, can someone enlighten me why this happens?
During training, i calculate the loss using all positive anchors with iou>0.4, negative anchors with iou<0.3. negative anchors are selected using hard negative mining, they are 3 times the number of positive anchors.
During testing, i use all initialized anchors to do a forward pass and do NMS on all anchors