About the activitynet features.

[sqiangcao99]

Hi,

For ActivityNet pre-trained model, which two configuration files are used for rgb and flow?

[xumingze0308]

We used the "clip_rgb" ones. You can also directly use the feature from TeSTra.

[sqiangcao99]

@xumingze0308. Hi, thanks for your help. Currently, I have some proplems reproducing the resulting the results(6% lower) on TVSeries using the ActivityNet pretrained features. Specifically,

1. I extract the rgb frames with resolution of short-side 320 and flows with resolution of 340X256;

2. The rgb frames and flow frams are preprocessed according to the config file of mmaction2, which is

   # RGB
   data_pipeline = [
       dict(type='RawFrameDecode'), 
       dict(type='CenterCrop', crop_size=256),
       dict(type='Normalize', **args.img_norm_cfg),
       dict(type='FormatShape', input_format='NCHW'),
       dict(type='Collect', keys=['imgs'], meta_keys=[]),
       dict(type='ToTensor', keys=['imgs']), 
   ]
   
   data_pipeline = [
       dict(type='RawFrameDecode'),
       dict(type='Resize', scale=(-1, 256)),
       dict(type='TenCrop', crop_size=224),
       dict(type='Normalize', **args.img_norm_cfg),
       dict(type='FormatShape', input_format='NCHW_Flow'),
       dict(type='Collect', keys=['imgs'], meta_keys=[]),
       dict(type='ToTensor', keys=['imgs'])
   ]

3. The extracting code is modified based on https://github.com/open-mmlab/mmaction2/blob/master/tools/data/activitynet/tsn_feature_extraction.py.

Is the whole process correct?

[xumingze0308]

For both rgb and flow, we don't conduct CenterCrop, but directly use the default resolution.