update

README.md

@@ -103,6 +103,6 @@ axmodel input: 640x480
 
 ```
 
-视频降噪样例（点击播放）：
+视频降噪样例：
 
-[![Video Denoising Demo](fastDVDnet/res/people-sunset_axmodel_side_by_side.mp4)](fastDVDnet/res/people-sunset_axmodel_side_by_side.mp4)
+![Video Denoising Demo](fastDVDnet/res/people-sunset_axmodel_side_by_side.gif)

fastDVDnet/python/axmodel_video_infer.py

@@ -103,8 +103,22 @@ def denoise_axmodel(frames_bgr, noise_sigma_01, sess, ax_h, ax_w,
 def write_side_by_side(frames_orig, frames_den, out_path, fps, label=True):
     h, w = frames_orig[0].shape[:2]
     os.makedirs(os.path.dirname(os.path.abspath(out_path)), exist_ok=True)
-    fourcc = cv2.VideoWriter_fourcc(*"mp4v")
+
+    ext = os.path.splitext(out_path)[1].lower()
+
+    if ext == '.gif':
+        _write_gif(frames_orig, frames_den, out_path, fps, w, h, label)
+    else:
+        _write_mp4(frames_orig, frames_den, out_path, fps, w, h, label)
+
+
+def _write_mp4(frames_orig, frames_den, out_path, fps, w, h, label):
+    fourcc = cv2.VideoWriter_fourcc(*"avc1")
     writer = cv2.VideoWriter(out_path, fourcc, fps, (w * 2, h))
+    if not writer.isOpened():
+        # fallback to mp4v if avc1 not available
+        fourcc = cv2.VideoWriter_fourcc(*"mp4v")
+        writer = cv2.VideoWriter(out_path, fourcc, fps, (w * 2, h))
     if not writer.isOpened():
         raise RuntimeError("video writer fail: {}".format(out_path))
     for orig, den in zip(frames_orig, frames_den):
@@ -118,6 +132,30 @@ def write_side_by_side(frames_orig, frames_den, out_path, fps, label=True):
     writer.release()
 
 
+def _write_gif(frames_orig, frames_den, out_path, fps, w, h, label):
+    try:
+        from PIL import Image
+    except ImportError:
+        raise RuntimeError("GIF output requires Pillow: pip install Pillow")
+
+    gif_frames = []
+    for orig, den in zip(frames_orig, frames_den):
+        canvas = np.concatenate([orig, den], axis=1)
+        if label:
+            cv2.putText(canvas, "Original", (16, 34), cv2.FONT_HERSHEY_SIMPLEX,
+                        1.0, (0, 255, 255), 2, cv2.LINE_AA)
+            cv2.putText(canvas, "AXModel Denoised", (w + 16, 34), cv2.FONT_HERSHEY_SIMPLEX,
+                        1.0, (0, 255, 255), 2, cv2.LINE_AA)
+        canvas_rgb = cv2.cvtColor(canvas, cv2.COLOR_BGR2RGB)
+        gif_frames.append(Image.fromarray(canvas_rgb))
+
+    duration = int(1000.0 / fps)
+    gif_frames[0].save(
+        out_path, save_all=True, append_images=gif_frames[1:],
+        duration=duration, loop=0, optimize=True,
+    )
+
+
 def main():
     parser = argparse.ArgumentParser(description="FastDVDnet axmodel video inference")
     parser.add_argument("--axmodel", type=str, default="./fastdvdnet_640x480.axmodel")
@@ -126,6 +164,8 @@ def main():
     parser.add_argument("--out_dir", type=str, default="./")
     parser.add_argument("--max_frames", type=int, default=0, help="0=all")
     parser.add_argument("--no_label", action="store_true")
+    parser.add_argument("--format", type=str, default="gif", choices=["mp4", "gif"],
+                        help="output format")
     args = parser.parse_args()
 
     sess = axe.InferenceSession(args.axmodel, providers=["AxEngineExecutionProvider"])
@@ -142,7 +182,7 @@ def main():
     print("outputs:", [(n, list(o.shape)) for n, o in zip(output_names, sess.get_outputs())])
 
     base = os.path.splitext(os.path.basename(args.video))[0]
-    out_path = os.path.join(args.out_dir, base + "_axmodel_side_by_side.mp4")
+    out_path = os.path.join(args.out_dir, base + "_axmodel_side_by_side." + args.format)
 
     t0 = time.time()
     frames_bgr, fps, orig_w, orig_h = read_video(args.video, args.max_frames)