pre-compute the features for hard example mining

[maojiaoli]

could you share me for pre-computeed the features date for hard example mining? my email is : 1808272971@qq.com thank you !

[pvoigtlaender]

I'm happy to share the data, but I'd first like to clarify what data exactly you need so I can share only what you actually need, since some of the data is quite big. I'll post this publicly here since others might also be interested in that, but you can then tell me what exactly you need.

I have the following data: -pre-computed nearest neighbor index structure (together for the frames in GOT10-k, LaSOT, YouTube-VOS, and ImageNet VID), this is ~9GB and easy to use, but not very flexible, since it's all 4 datasets together and you can't add or remove other data (for my own reference: /globalwork/voigtlaender/data/hard_example_mining_index/index_all/index.ann, /globalwork/voigtlaender/data/hard_example_mining_index/index_all/names.txt) -the pre-computed embeddings for the 4 datasets which were used to generate the data structure (for my own reference: /globalwork/voigtlaender/data/hard_example_mining_index/${DATASET}/triplet_feats/), sizes: GOT10k 1.9GB, ImageNet VID: 2.5GB, LaSOT: 3.6 GB, YouTube-VOS: 0.2 GB -the pre-computed RoI aligned backbone features for the ground truth bounding boxes (you will only need this if you want to train Siam R-CNN with hard example mining, and not if you want to train any other method), this is by far the biggest data, didn't compute the exact size for all datasets yet, YouTubeVOS is 11GB (for my own reference: /globalwork/voigtlaender/data/hard_example_mining_index/${DATASET}/det_feats_compressed/) -images cropped to the objects for visualization purposes (probably not very important since you can easily generate it yourself, for my own reference: /globalwork/voigtlaender/data/hard_example_mining_index/${DATASET}/crops)