File size: 8,259 Bytes
692e2af
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
import os
import shutil
from tqdm import tqdm
import cv2
import pandas as pd
import torch
from PIL import Image
from transformers import Owlv2Processor, Owlv2ForObjectDetection
import math
import zipfile
from utils import plot_predictions, mp4_to_png, vid_stitcher

def owl_batch_video(

    input_vids: list[str],

    target_prompt: list[str],

    species_prompt: str,

    threshold: float,

    fps_processed: int = 1,

    scaling_factor: float = 0.5, 

    batch_size: int = 8, 

    save_dir: str = "temp/"

):
    pos_preds = []
    neg_preds = []

    df = pd.DataFrame(columns=["video path", "detection?"])

    for vid in input_vids:
        detection = owl_video_detection(vid, 
                            target_prompt, 
                            species_prompt, 
                            threshold, 
                            fps_processed=fps_processed,
                            scaling_factor=scaling_factor,
                            batch_size=batch_size,
                            save_dir=save_dir)

        if detection == True:
            pos_preds.append(vid)
            row = pd.DataFrame({"video path": [vid], "detection?": ["True"]})
            df = pd.concat([df, row], ignore_index=True)
        else:
            neg_preds.append(vid)
            row = pd.DataFrame({"video path": [vid], "detection?": ["False"]})
            df = pd.concat([df, row], ignore_index=True)

    # save the df
    df.to_csv(save_dir + "detection_results.csv")
    
    # zip the save_dir
    zip_file = f"{save_dir}/results.zip"
    zip_directory(save_dir, zip_file)

    return zip_file


    
def zip_directory(folder_path, output_zip_path):
    with zipfile.ZipFile(output_zip_path, 'w', zipfile.ZIP_DEFLATED) as zipf:
        for root, dirs, files in os.walk(folder_path):
            for file in files:
                file_path = os.path.join(root, file)
                # Write the file with a relative path to preserve folder structure
                arcname = os.path.relpath(file_path, start=folder_path)
                zipf.write(file_path, arcname)


def preprocess_text(text_prompt: str, num_prompts: int = 1):
    """

    Takes a string of text prompts and returns a list of lists of text prompts for each image. 

    i.e. text_prompt = "a, b, c" -> [["a", "b", "c"], ["a", "b", "c"]]

    """
    text_prompt = [s.strip() for s in text_prompt.split(",")]
    text_queries = [text_prompt] * num_prompts
    # print("text_queries:", text_queries)
    return text_queries

def owl_batch_prediction(

        images: torch.Tensor,

        text_queries : list[str], # assuming that every image is queried with the same text prompt

        threshold: float,

        processor, 

        model,

        device: str = 'cuda'

    ):

    inputs = processor(text=text_queries, images=images, return_tensors="pt").to(device)
    with torch.no_grad():
        outputs = model(**inputs)

     # Target image sizes (height, width) to rescale box predictions [batch_size, 2]
    target_sizes = torch.Tensor([img.size[::-1] for img in images]).to(device)
    # Convert outputs (bounding boxes and class logits) to COCO API, resizes to original image size and filter by threshold
    results = processor.post_process_object_detection(outputs=outputs, target_sizes=target_sizes, threshold=threshold)

    return results


def count_pos(phrases: list[str], text_targets: list[str]) -> int:
    """

    Counts how many phrases in the list match any of the target phrases.



    Args:

        phrases: A list of strings to evaluate.

        text_targets: A list of target strings to match against.



    Returns:

        The number of phrases that match any of the targets.

    """
    if len(phrases) == 0 or len(text_targets) == 0:
        return 0
    target_set = set(text_targets)
    return sum(1 for phrase in phrases if phrase in target_set)


def owl_video_detection(

        vid_path: str,

        text_target: list[str], 

        text_prompt: str,

        threshold: float,

        fps_processed: int = 1,

        scaling_factor: float = 0.5,

        processor = Owlv2Processor.from_pretrained("google/owlv2-base-patch16-ensemble"),

        model = Owlv2ForObjectDetection.from_pretrained("google/owlv2-base-patch16-ensemble").to('cuda'),

        device: str = 'cuda',

        batch_size: int = 8,

        save_dir: str = "temp/",

        ): 
    """

    Runs owl on a video and saves the results to a dataframe. 

    Returns True if text_target is detected in the video, False otherwise. 

    Stops running owl when a text_target is detected.

    """
    os.makedirs(save_dir, exist_ok=True)
    os.makedirs(f"{save_dir}/positives", exist_ok=True)
    os.makedirs(f"{save_dir}/negatives", exist_ok=True)

    # set up df for results
    df  = pd.DataFrame(columns=["frame", "boxes", "scores", "labels", "count"])

    # create new dirs and paths for results
    filename = os.path.splitext(os.path.basename(vid_path))[0]
    frames_dir = f"{save_dir}/{filename}_frames"
    os.makedirs(frames_dir, exist_ok=True)

    # process video and create a directory of video frames
    fps = mp4_to_png(vid_path, frames_dir, scaling_factor)

    # get all frame paths
    frame_filenames = os.listdir(frames_dir)

    frame_paths = []  # list of frame paths to process based on fps_processed
    # for every frame processed, add to frame_paths
    for i, frame in enumerate(frame_filenames):
        if i % fps_processed == 0:
            frame_paths.append(os.path.join(frames_dir, frame))

    # run owl in batches
    for i in tqdm(range(0, len(frame_paths), batch_size), desc="Running batches"):
        frame_nums = [i*fps_processed for i in range(batch_size)]
        batch_paths = frame_paths[i:i+batch_size]  # paths for this batch 
        images = [Image.open(image_path) for image_path in batch_paths]

        # run owl on this batch of frames
        text_queries = preprocess_text(text_prompt, len(batch_paths))
        results = owl_batch_prediction(images, text_queries, threshold, processor, model, device)

        # get the boxes, logits, and phrases for this batch
        label_ids = []
        for entry in results:
            if entry['labels'].numel() > 0:
                label_ids.append(entry['labels'].tolist())
            else:
                label_ids.append(None)

        text = text_queries[0] # assuming that all texts in query are the same for each image
        labels = []
        # convert label_ids to phrases, if no phrases, append None
        for idx in label_ids:
            if idx is not None:
                idx = [text[id] for id in idx]
                labels.append(idx)
            else: 
                labels.append([])
        
        batch_pos = 0
        for j, image in enumerate(batch_paths):
            boxes = results[j]['boxes'].cpu().numpy()
            scores = results[j]['scores'].cpu().numpy()
            print(labels[j], text_target, count_pos(labels[j], text_target))
            count = count_pos(labels[j], text_target)
            row = pd.DataFrame({"frame": [image], "boxes": [boxes], "scores": [scores], "labels": [labels[j]], "count": count})
            df = pd.concat([df, row], ignore_index=True)
            
            # if there are detections, save the frame replacing the original frame
            if count > 0:
                annotated_frame = plot_predictions(image, labels[j], scores, boxes)
                cv2.imwrite(image, annotated_frame)
                batch_pos += 1
    
        # if more than 2/3 batch frames are positive, return True
        if batch_pos > math.ceil(2/3*batch_size):
            vid_stitcher(frames_dir, f"{save_dir}/positives/{filename}_{threshold}.mp4", fps)
            shutil.rmtree(frames_dir) # delete the frames to save space
            df.to_csv(f"{save_dir}/positives/{filename}_{threshold}.csv", index=False)
            return True
    
    shutil.rmtree(frames_dir) # delete the frames to save space
    df.to_csv(f"{save_dir}/negatives/{filename}_{threshold}.csv", index=False)
    return False