anno_preprocess/preprocess_perfeature_v2.py

##################
#To detect wearer speaker
##################

import json
import pickle

import numpy as np

# Error : somewhere [0,0,0] produced - SOLVED

av_train_annotation_path =  '/home/junhyeok/projects/turn-taking-dataset/ego4d_dataset/v2/annotations/av_train.json'

with open(av_train_annotation_path, "r") as f:
    av_train_annotations = json.load(f)

av_train_processed = {}

for video in av_train_annotations['videos'] :
    for clip in video['clips'] :
        uid = clip['clip_uid']
        av_train_processed[uid] = {}
        # Make embedding template
        end_frame = clip['clip_end_frame']
        #feature_num = int(np.ceil(end_frame/6))
        feature_num = int(np.round(end_frame/6))
        if feature_num != 1500  :
            print('uid : {} , feature_length : {}'.format(uid , feature_num))
        anno = np.zeros((feature_num , 3))
        # Suppose no action in frame for default
        anno[:,0]=1

        i=0
        for transcript in clip['transcriptions']:

            # To make annotation by clip frame, need to subtract video_start_frame from transcription's start_frame
            if i==0 :
                initial_frame = clip['video_start_frame']
                i+=1

            #Save frame range
            start_frame = transcript['video_start_frame'] - initial_frame
            end_frame = transcript['video_end_frame'] - initial_frame

            #Save action encoding
            if int(transcript['person_id']) == 0 :
                encode_num = 1
            if int(transcript['person_id']) >= 1 :
                encode_num = 2
            if int(transcript['person_id']) == -1 :
                encode_num = 0
            
            #Save encoding by feature
            start = start_frame //6 -1
            end = end_frame //6
            if start <0 :
                start = 0
            #Offset 5 for wearer
            if encode_num == 1 :
                start_ = start-5
                if start_<0 :
                    start_=0
                if end - start == 0 :
                    anno[start_:start+1 , 0] = 0
                    anno[start_:start+1 , encode_num]=1    
                else :
                    anno[start_ : end , 0] = 0
                    anno[start_ : end , encode_num]=1
            #For background , normal speaker
            else:
                if end-start == 1 :
                    anno[start , 0] = 0
                    anno[start , encode_num]=1    
                else :
                    anno[start:end , 0] = 0
                    anno[start:end , encode_num]=1
                

            #Save annotation to uid list
            av_train_processed[uid]['anno'] = anno
            av_train_processed[uid]['feature_length'] = feature_num

print('Processing DONE!')

with open('train_perfeature_test.pickle' , 'wb') as f :
    pickle.dump(av_train_processed , f)

print('SAVED')