he010103
[Track2: ReID] Question about 9 pairs of features, combination of losses, and some missing codes.
Dear Zhiqun He,
I am getting a lot of help for you providing the code. I have some questions about that code.
- Could you tell me how to extract 9 pairs of features for query and gallery in the provided google drive?
Through the codes of run.sh and submit.py, I can get a pair of features for query and test (gallery). However, in the gen_track2.py code, the features of the paired query and test (gallery) I learned are not loaded. To be more specific, 9 pairs of pre-computed (downloadable) features for queries and tests (gallery) are loaded, and the final result is output by these feature vectors. In other words, the gen_track2.py code provides for testing pre-trained features (9 pairs) rather than using features I trained. I know that I can download these feature vectors in your google drive, but I don't think there is a way to extract these 9 pairs of features in the code and workshop paper. I wonder how to extract 9 pairs of features. Are they about changing losses? or Are they about changing training datasets? Could you tell me how to extract them?
- There are several options for the combination of loss in the code. Could you tell which loss combination has the best performance?
When you submit to the challenge, I want to know if you use the loss combination below. (if cfg.MODEL.IF_WITH_CENTER == 'yes' -> centroid mode) cfg.MODEL.METRIC_LOSS_TYPE =?
- 'center' -> CE + center
- 'range_center' -> CE + center + range
- 'triplet_center' -> CE + triplet + track_triplet + center
- 'triplet_range_center' -> CE + triplet + track_triplet + center + range
- The code for "Reranking with spatial-temporal cue" seems to be missing. Any plans to release this code?
Looking at the gen_track2.py code, there is the following code. "from metrics import track_reranking_weight_feat" However, I haven't found any metrics folder or track_reranking_weight_feat function anywhere. Isn't that function provided?
Thanks Seokeon Choi
