Position-Focused-Attention-Network
Position-Focused-Attention-Network copied to clipboard
HaoYang0123
About the experiment results on MS COCO
(1) Do you evaluate your model in testall dataset as SCAN with 1K images (5-folds) and 5K images? Can you provide the box information of testall dataset? (2) It seems that PFAN i-t is not as effective as PFAN t-i after comparing PFAN i-t with SCAN i-t. Do you have any idea about this?
- We haven't test the 5-folds results. And so we don't have the box information of testall dataset. This will be considered in our future work.
- The performance of PFAN i-t is not as good as PFAN t-i. The method based on "t-i" compute the weighted word features using attended on image region features, then compute the similarity between the original word features and weighted word features. It may be better based on words than "i-t" based on regions because the region features are extracted by Faster R-CNN. The region features may not be so good, for example, not all of 36 regions are important when computing similarity.
Thanks! Another thing. I used my own preprocessing code for the box information (eg. getting region index, computing the percentage of the overlapping area, normalizing the overlapping weights) and your model code. However, I failed to replicate your results on MS COCO. Can you provide your code for preprocessing the box information?
Recently I am working other projects and will upload the preprocessing box information code later.
@miracle24 @HaoYang0123 Hi, sorry to bother you. Can I ask you some questions about reproducing results on COCO? I download the "coco_precomp" zip file and get the boxes' information of images. But there are only 1000 images for validation and test data that have boxes, while the validation and test set includes 5000 images in total. Here I got an error and I don't know if I should change something somewhere.
Thanks so much and look forward to your reply.
@miracle24 @HaoYang0123 Hi, sorry to bother you. Can I ask you some questions about reproducing results on COCO? I download the "coco_precomp" zip file and get the boxes' information of images. But there are only 1000 images for validation and test data that have boxes, while the validation and test set includes 5000 images in total. Here I got an error and I don't know if I should change something somewhere.
Thanks so much and look forward to your reply.
@Sammy42779 同学请问您咋解决的，我也遇到这个问题了
