Hello, thank you very much for your great works.
I am trying to replicate your model at detection on AVA Dataset for feature extraction part in my work. My goal is to extract the features of every bounding box.
However, after successfully making the code for extracting, I observed that many bounding boxes have zeros only features.
I have rescaled my bounding boxes together with frames by your datasets.transform.random_short_side_scale_jitter.
I think the problem lies at the parameter cfg.DETECTION.SPATIAL_SCALE_FACTOR which is set to 16 as default, but I think it should be set to 32 because the original short side of frame is 256 and should become 8 before being fed to RoIAlign.
If my thought is right. I would like to ask you how to replicate the settings of your Detection model as the paper said: "We set the spatial stride of res_5 to 1 (instead of 2), and use a dilation of 2 for its filters. This increases the spatial resolution of res_5 by 2x." ?
Hello, thank you very much for your great works.
I am trying to replicate your model at detection on AVA Dataset for feature extraction part in my work. My goal is to extract the features of every bounding box.
However, after successfully making the code for extracting, I observed that many bounding boxes have zeros only features.
I have rescaled my bounding boxes together with frames by your
datasets.transform.random_short_side_scale_jitter.I think the problem lies at the parameter
cfg.DETECTION.SPATIAL_SCALE_FACTORwhich is set to 16 as default, but I think it should be set to 32 because the original short side of frame is 256 and should become 8 before being fed to RoIAlign.If my thought is right. I would like to ask you how to replicate the settings of your Detection model as the paper said: "We set the spatial stride of res_5 to 1 (instead of 2), and use a dilation of 2 for its filters. This increases the spatial resolution of res_5 by 2x." ?