Datasets:

Modalities:
Video
Audio
Languages:
English
ArXiv:
License:
File size: 7,480 Bytes
f89df01
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
import os
import json
import torch
import numpy as np
from tqdm import tqdm   
from torch import Tensor

from typing import List, Dict, Set


from evaluation.utils import heatmap_interpolation, video_interpolation, \
    convert_ann_to_mask


task_dict = {
    "single": "Single-Sound",
    "mixed": "Mixed-Sound",
    "multi": "Multi-Entity",
    "off-screen": "Off-Screen"
}

class Evaluator:
    def __init__(self, data_root: str, model: str, benchmark_path: str, 
                heatmap_thresholds: List=[0.1, 0.15, 0.2], height: int=360, width: int=640, 
                frame_sample_rate: int=8, video: bool=True):
        
        # data path
        self.data_root = data_root
        self.model = model
        self.benchmark_path = benchmark_path
        
        self.heatmap_thresholds = heatmap_thresholds
        self.height = height
        self.width = width
        self.frame_sample_rate = frame_sample_rate
        
        self.video = video
        
        self.threshold_results = json.load(open(os.path.join(self.data_root, self.model, f"heatmap_threshold.json"), "r"))
        self.evaluation_results = {} # {frame_id: {heatmap_percent: {metric: value}, ...}}


    def compute(self):
        data_path = os.path.join(self.data_root, self.model, "heatmap")

        print(f"Evaluating {'Video' if self.video else 'Image'} Heatmap")
        self.evaluate_video(data_path) if self.video else self.evaluate_image(data_path)

        self.save_results()


    def evaluate_image(self, data_path: str):
        
        for data_file in tqdm(os.listdir(data_path)):
            if not data_file.endswith(".npy"):
                continue
            
            frame_num = int(data_file.split("_")[-1].split(".")[0])
            video_id = "_".join(data_file.split("_")[:-1])
            metadata_path = os.path.join(self.benchmark_path, video_id, f"{frame_num:05d}.json")

            infer = torch.tensor(np.load(os.path.join(data_path, data_file))).unsqueeze(0).unsqueeze(0)
            infer = heatmap_interpolation(infer, self.height, self.width)

            metadata = json.load(open(metadata_path, "r"))
            self.evaluate(video_id, frame_num, infer, metadata["annotations"])


    def evaluate_video(self, data_path: str):

        for data_file in tqdm(os.listdir(data_path)):
            if not data_file.endswith(".npy"):
                continue
            
            video_id = data_file.split(".")[0]
            metadata_path = os.path.join(self.benchmark_path, video_id)
            
            infer = torch.tensor(np.load(os.path.join(data_path, data_file)))
            infer = video_interpolation(infer, self.frame_sample_rate)
            
            for frame_data in os.listdir(metadata_path):
                if frame_data.endswith(".jpg"):
                    continue

                frame_num = int(frame_data.split(".")[0])
                if frame_num >= infer.shape[0]:
                    print(f"Frame number is larger than infer shape: {frame_num} >= {infer.shape[0]}")
                
                infer_map = infer[frame_num].unsqueeze(0)
                infer_map = heatmap_interpolation(infer_map, self.height, self.width)

                metadata = json.load(open(os.path.join(metadata_path, frame_data), "r"))
                self.evaluate(video_id, frame_num, infer_map, metadata["annotations"])
                


    def evaluate(self, video_id: str, frame_id: int, infer_map: np.ndarray, annotations: Dict):
        heatmap_dict = dict() # {percent: map}
        for percent_thr in self.heatmap_thresholds:
            thr = self.threshold_results[str(percent_thr)]
            heatmap_dict[percent_thr] = (infer_map > thr)
        
        accumulated_mask = np.zeros((self.height, self.width))
        off_screen = False
        
        pixel_statistics = {thr: {"tp": 0, "fp": 0, "fn": 0} for thr in self.heatmap_thresholds}
        instance_statistics = {thr: {
            "total_tp": 0, "total_fn": 0,
            "single_tp": 0, "single_fn": 0,
            "mixed_tp": 0, "mixed_fn": 0,
            "multi_tp": 0, "multi_fn": 0
            } for thr in self.heatmap_thresholds}
        cious = {thr: 0 for thr in self.heatmap_thresholds}
        
        # Instance-Level Evaluation
        for ann in annotations:
            if ann["task"] == task_dict["off-screen"]:
                off_screen = True
                continue
            mask = convert_ann_to_mask(ann, self.height, self.width)
            accumulated_mask = np.logical_or(accumulated_mask, mask)
            
            for percent_thr in self.heatmap_thresholds:
                heatmap = heatmap_dict[percent_thr]
                instance_statistics[percent_thr] = self.evaluate_instance_predictions(mask, heatmap, ann["task"], instance_statistics[percent_thr])
        
        # Pixel-Level Evaluation
        for percent_thr in self.heatmap_thresholds:
            heatmap = heatmap_dict[percent_thr]
            pixel_statistics[percent_thr] = self.evaluate_pixel_predictions(accumulated_mask, heatmap, pixel_statistics[percent_thr])
            
            if not off_screen:
                cious[percent_thr] = self.evaluate_cIoU(accumulated_mask, heatmap)

        self.evaluation_results[f"{video_id}_{frame_id}"] = {
            "pixel_statistics": pixel_statistics,
            "instance_statistics": instance_statistics,
            "off_screen": off_screen,
            "cious": cious
        }


    def evaluate_instance_predictions(self, mask: np.ndarray, heatmap: np.ndarray, task: str, instance_statistics: Dict):
        if np.sum(np.logical_and(mask, heatmap)) / np.sum(mask) >= self.instance_threshold:
            instance_statistics["total_tp"] += 1
            if task == task_dict["single"]:
                instance_statistics["single_tp"] += 1
            elif task == task_dict["mixed"]:
                instance_statistics["mixed_tp"] += 1
            elif task == task_dict["multi"]:
                instance_statistics["multi_tp"] += 1
        else:
            instance_statistics["total_fn"] += 1
            if task == task_dict["single"]:
                instance_statistics["single_fn"] += 1
            elif task == task_dict["mixed"]:
                instance_statistics["mixed_fn"] += 1
            elif task == task_dict["multi"]:
                instance_statistics["multi_fn"] += 1
        return instance_statistics
        
    
    def evaluate_pixel_predictions(self, acc_mask: np.ndarray, heatmap: np.ndarray, pixel_statistics: Dict):
        tp_pixel = int(np.sum(np.logical_and(acc_mask, heatmap)))
        fp_pixel = int(np.sum(heatmap) - tp_pixel)
        fn_pixel = int(np.sum(acc_mask) - tp_pixel)
        
        pixel_statistics["tp"] = tp_pixel
        pixel_statistics["fp"] = fp_pixel
        pixel_statistics["fn"] = fn_pixel
        return pixel_statistics
        
        
    def evaluate_cIoU(self, acc_mask: np.ndarray, heatmap: np.ndarray):
        ciou = np.sum(heatmap * acc_mask) / (np.sum(acc_mask) + np.sum(heatmap * (acc_mask == 0)))
        return ciou
    
    
    def save_results(self):
        save_folder = os.path.join(self.data_root, self.model)
        os.makedirs(save_folder, exist_ok=True)
        
        save_path = os.path.join(save_folder, f"evaluation_results.json")
        with open(save_path, "w") as f:
            json.dump(self.evaluation_results, f)
            
        print(f"Results are saved at {save_path}")