update video rotation when doing detection

detect.py

@@ -61,7 +61,7 @@ def detect(save_img=False):
     t0 = time.time()
     img = torch.zeros((1, 3, imgsz, imgsz), device=device)  # init img
     _ = model(img.half() if half else img) if device.type != 'cpu' else None  # run once
-    for path, img, im0s, vid_cap in dataset:
+    for path, img, im0s, vid_cap, rotation in dataset:
         img = torch.from_numpy(img).to(device)
         img = img.half() if half else img.float()  # uint8 to fp16/32
         img /= 255.0  # 0 - 255 to 0.0 - 1.0
@@ -133,8 +133,8 @@ def detect(save_img=False):
 
                         fourcc = 'mp4v'  # output video codec
                         fps = vid_cap.get(cv2.CAP_PROP_FPS)
-                        w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH))
-                        h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+                        w = int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH)) if not rotation else int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
+                        h = int(vid_cap.get(cv2.CAP_PROP_FRAME_HEIGHT)) if not rotation else int(vid_cap.get(cv2.CAP_PROP_FRAME_WIDTH))
                         vid_writer = cv2.VideoWriter(save_path, cv2.VideoWriter_fourcc(*fourcc), fps, (w, h))
                     vid_writer.write(im0)
 

requirements.txt

@@ -12,6 +12,8 @@ tensorboard>=2.2
 torch>=1.6.0
 torchvision>=0.7.0
 tqdm>=4.41.0
+scikit-video
+ffmpeg
 
 # logging -------------------------------------
 # wandb

utils/datasets.py

@@ -8,6 +8,7 @@
 
 import cv2
 import math
+import skvideo.io
 import numpy as np
 import torch
 from PIL import Image, ExifTags
@@ -122,9 +123,19 @@ def __init__(self, path, img_size=640):
         images = [x for x in files if os.path.splitext(x)[-1].lower() in img_formats]
         videos = [x for x in files if os.path.splitext(x)[-1].lower() in vid_formats]
         ni, nv = len(images), len(videos)
+        videos_rotation = [None for _ in videos]
+        for index in range(nv):
+            metadata = skvideo.io.ffprobe(videos[index])
+            if 'video' in metadata and 'tag' in metadata['video']:
+                tags = metadata['video']['tag']
+
+                for tag in tags:
+                    if tag['@key'] == 'rotate':
+                        videos_rotation[index] = tag['@value']
 
         self.img_size = img_size
         self.files = images + videos
+        self.rotation = [None for _ in images] + videos_rotation
         self.nf = ni + nv  # number of files
         self.video_flag = [False] * ni + [True] * nv
         self.mode = 'images'
@@ -143,6 +154,7 @@ def __next__(self):
         if self.count == self.nf:
             raise StopIteration
         path = self.files[self.count]
+        rotation = self.rotation[self.count]
 
         if self.video_flag[self.count]:
             # Read video
@@ -168,6 +180,10 @@ def __next__(self):
             assert img0 is not None, 'Image Not Found ' + path
             print('image %g/%g %s: ' % (self.count, self.nf, path), end='')
 
+        # Rotation Valid
+        if rotation:
+            img0 = cv2.rotate(img0, 0)
+
         # Padded resize
         img = letterbox(img0, new_shape=self.img_size)[0]
 
@@ -176,7 +192,7 @@ def __next__(self):
         img = np.ascontiguousarray(img)
 
         # cv2.imwrite(path + '.letterbox.jpg', 255 * img.transpose((1, 2, 0))[:, :, ::-1])  # save letterbox image
-        return path, img, img0, self.cap
+        return path, img, img0, self.cap, rotation
 
     def new_video(self, path):
         self.frame = 0
@@ -246,7 +262,7 @@ def __next__(self):
         img = img[:, :, ::-1].transpose(2, 0, 1)  # BGR to RGB, to 3x416x416
         img = np.ascontiguousarray(img)
 
-        return img_path, img, img0, None
+        return img_path, img, img0, None, None
 
     def __len__(self):
         return 0
@@ -319,7 +335,7 @@ def __next__(self):
         img = img[:, :, :, ::-1].transpose(0, 3, 1, 2)  # BGR to RGB, to bsx3x416x416
         img = np.ascontiguousarray(img)
 
-        return self.sources, img, img0, None
+        return self.sources, img, img0, None, None
 
     def __len__(self):
         return 0  # 1E12 frames = 32 streams at 30 FPS for 30 years