Add the modified openpose and add confidence level filtering.

README.md

@@ -16,11 +16,11 @@ The entire project is divided into two parts: `text_or_gloss2pose` and `pose2vid
 first, change `good morning afternoon`, and run:
 
 ```python
-python text2pose.py --gloss "good morning afternoon" --scale-y 1.3 --scale-x 1.2 --npz-interpolation 20 --width 512 --height 768 --fps 30 --hide-torso-lines true --ref-image-path ref_img_dict/01.jpeg
+python text2pose.py --gloss "good morning afternoon" --scale-y 1.4 --scale-x 1.3 --npz-interpolation 20 --width 512 --height 768 --fps 30 --hide-torso-lines true --ref-image-path ref_img_dict/01.jpeg
 
 or
 
-python text2pose.py --gloss "where is my cat" --scale-y 1.4 --scale-x 1.3 --npz-interpolation 20 --width 512 --height 768 --fps 30 --hide-torso-lines false
+python text2pose.py --gloss "where is my cat" --scale-y 1.5 --scale-x 1.5 --npz-interpolation 20 --width 512 --height 512 --fps 30 --hide-torso-lines false
 ```
 
 second, change `generated_pose_video/GOOD_MORNING_AFTERNOON_25073_36863_01435_20250804_160832.mp4` & `ref_img_dict/03.jpeg`, and run:

README_v1.md

@@ -0,0 +1,199 @@
+# Intro
+
+The entire project is divided into two parts: `text_or_gloss2pose` and `pose2video`. "Pose" here refers to the skeleton pose video after drawing.
+
+- In the text2pose section (`Expected 4 seconds / one video`), I first used the DWpose tool to extract the npz format pose information for each word video frame by frame, and then merged each frame's npz into a large npz file. That is to say, we know the consecutive poses corresponding to each word. These consecutive poses are stored in the merged npz file. (We can also use deep learning models to generate these poses, but for commercial-level accuracy, this original data Dict method will be much more accurate.) 
+
+- In the pose2video section (`Expected 6 minutes / one video`), we render the visualized poses of the skeletons generated by each word. I used the ControlNext-SVD model for this. 
+
+# For Users
+
+1. cd StableSigner
+2. conda env create -f env/environment02_controlnext.yml
+3. run `conda activate controlnext`
+4. run `pip install -r env/requirements02_controlnext.txt`
+
+first, change `good morning afternoon`, and run:
+
+```python
+python text2pose.py --gloss "good morning afternoon" --scale-y 1.3 --scale-x 1.2 --npz-interpolation 20 --width 512 --height 768 --fps 30 --hide-torso-lines true --ref-image-path ref_img_dict/01.jpeg
+
+or
+
+python text2pose.py --gloss "where is my cat" --scale-y 1.4 --scale-x 1.3 --npz-interpolation 20 --width 512 --height 768 --fps 30 --hide-torso-lines false
+```
+
+second, change `generated_pose_video/GOOD_MORNING_AFTERNOON_25073_36863_01435_20250804_160832.mp4` & `ref_img_dict/03.jpeg`, and run:
+
+```python
+CUDA_VISIBLE_DEVICES=0 python ControlNeXt-SVD-v2/run_controlnext_optimized.py \
+  --pretrained_model_name_or_path stabilityai/stable-video-diffusion-img2vid-xt-1-1 \
+  --output_dir generated_sign_video \
+  --max_frame_num 240 \
+  --guidance_scale 3 \
+  --batch_frames 24 \
+  --sample_stride 2 \
+  --overlap 4 \
+  --height 768 \
+  --width 512 \
+  --controlnext_path ControlNeXt-SVD-v2/pretrained/controlnet.bin \
+  --unet_path ControlNeXt-SVD-v2/pretrained/unet.bin \
+  --pose_video_path generated_pose_video/GOOD_MORNING_AFTERNOON_25073_36863_01435_20250804_160832.mp4 \
+  --ref_image_path ref_img_dict/03.jpeg \
+  --precision fp16 \
+  --decode_chunk_size 2 \
+  --enable_geglu_optimization \
+  --enable_xformers 
+```
+
+**Note:** In fact, I haven't installed xformer and haven't conducted a comprehensive test of it. If not installed, it will not affect the normal operation.
+
+**Note:** It seems that the first time using pose2vid will result in poor quality and speed, and some kind of warm-up is necessary.
+
+**Note:** I have tried running two stages with a resolution of `512*512`, but it seems that it causes instability in the background.
+
+**Note:** In some cases, `512w*768h` can cause abnormalities in the torso. In rare instances, the torso may be mistaken for a hand. This issue should be able to be resolved by hiding the torso.
+
+**BUG:** For some unknown reason, the current robustness is not as good as before (24.12 of the demos I tried were better than the current pose2video). The background is relatively prone to change, and the bottom part of the upper body is prone to become blurry. I have taken two or three improvements to solve this problem, but so far, it seems to still occur strangely and persistently. If this phenomenon continues, I am considering directly using Docker in 24.12 as a replacement. (Later, I discovered that simply using Docker did not lead to any significant improvement: Just the background and the hands became a little more stable. I felt that I needed to try to find the most stable configuration.)
+
+
+## Demo Gallery
+
+<details>
+<summary>🎬 Click to view generated examples</summary>
+
+### Example 1: "good morning afternoon"
+
+<div align="center">
+    <h4>Pose Video & Final Video</h4>
+    <video width="80%" controls>
+        <source src="https://huggingface.co/datasets/FangSen9000/StableSigner/resolve/main/generated_sign_video/GOOD_MORNING_AFTERNOON_25073_36863_01435_20250804_160832_combined.mp4" type="video/mp4">
+        Your browser does not support the video tag.
+    </video>
+</div>
+
+---
+
+### Example 2: "hello world"
+
+<div align="center">
+    <h4>Pose Video & Final Video</h4>
+    <video width="80%" controls>
+        <source src="https://huggingface.co/datasets/FangSen9000/StableSigner/resolve/main/generated_sign_video/HELLO_WORLD_27184_63831_20250804_163910_combined.mp4" type="video/mp4">
+        Your browser does not support the video tag.
+    </video>
+</div>
+
+---
+
+### Example 3: "I will attend this dinner party"
+
+<div align="center">
+    <h4>Pose Video & Final Video</h4>
+    <video width="80%" controls>
+        <source src="https://huggingface.co/datasets/FangSen9000/StableSigner/resolve/main/generated_sign_video/I_WILL_ATTEND_THIS_DINNER_PARTY_28794_63364_04007_66628_16327_41325_20250804_164010_combined.mp4" type="video/mp4">
+        Your browser does not support the video tag.
+    </video>
+</div>
+
+---
+
+### Example 4: "not find yet"
+
+<div align="center">
+    <h4>Pose Video & Final Video</h4>
+    <video width="80%" controls>
+        <source src="https://huggingface.co/datasets/FangSen9000/StableSigner/resolve/main/generated_sign_video/NOT_FIND_YET_38868_21853_18535_20250804_164149_combined.mp4" type="video/mp4">
+        Your browser does not support the video tag.
+    </video>
+</div>
+
+---
+
+### Example 5: "where is my cat"
+
+<div align="center">
+    <h4>Pose Video & Final Video</h4>
+    <video width="80%" controls>
+        <source src="https://huggingface.co/datasets/FangSen9000/StableSigner/resolve/main/generated_sign_video/WHERE_IS_MY_CAT_63084_57261_37469_65313_20250804_164233_combined.mp4" type="video/mp4">
+        Your browser does not support the video tag.
+    </video>
+</div>
+
+### Note on video types:
+- **Pose video**: Shows the skeleton/pose generated from input text/gloss
+- **Final video**: The rendered sign language video with a virtual signer
+- **Combined view**: Side-by-side comparison of pose and final video
+
+</details>
+<br>
+
+# For developers
+
+1. git clone https://huggingface.co/datasets/FangSen9000/StableSigner
+2. cd StableSigner
+
+## tutorial00 Use DWpose to preprocess the original sign language video
+
+`tools-new-2025` is a tool that can extract the original video into a npz file containing dwpose skeleton posture information (a type of skeleton posture that is almost compatible with all current pose2video models).
+
+1. conda env create -f env/environment01_dwpose.yml
+2. run `conda activate dwpose`
+3. run `pip install -r env/requirements01_dwpose.txt`
+4. put raw sign viseos, like `tools-new-2025/input/test_dataset`
+5. cd `StableSigner/tools-new-2025/main`
+6. run `python pipeline00_extract_split_video_to_image.py` (Default frame rate is 30fps, and the video will be automatically cropped to a square shape.)
+7. run `python pipeline01_extract_dwpose_from_video.py` (It will generate files other than npz. We usually don't need them and can comment out the relevant code.)
+8. run `conda deactivate`, we finish the preprocess.
+
+## tutorial01 prepare pose dict
+
+1. cd StableSigner
+2. run `python utils/pipeline01_merge_frame_level_npz_from_video_dir.py tools-new-2025/output/test_dataset_results` 
+3. Place the corresponding "gloss-videoID" mapping in the pose dict
+
+## tutorial02 get pose video from text input
+
+```python
+
+# Default
+python text2pose.py --gloss "good morning afternoon" --npz-interpolation 20 --width 512 --height 768 --fps 30
+
+# Single gloss
+python text2pose.py --gloss "hello"
+
+# Multiple glosses with smoothing
+python text2pose.py --gloss "hello world" --smoothing fade --smoothing-frames 5
+
+# Custom video parameters
+python text2pose.py --gloss "good afternoon" --width 512 --height 768 --fps 30
+```
+
+## tutorial03 render pose video
+
+1. cd StableSigner
+2. conda env create -f env/environment02_controlnext.yml
+3. run `conda activate controlnext`
+4. run `pip install -r env/requirements02_controlnext.txt`
+
+```python
+CUDA_VISIBLE_DEVICES=0 python ControlNeXt-SVD-v2/run_controlnext_optimized.py \
+  --pretrained_model_name_or_path stabilityai/stable-video-diffusion-img2vid-xt-1-1 \
+  --output_dir generated_sign_video \
+  --max_frame_num 240 \
+  --guidance_scale 3 \
+  --batch_frames 24 \
+  --sample_stride 2 \
+  --overlap 4 \
+  --height 768 \
+  --width 512 \
+  --controlnext_path ControlNeXt-SVD-v2/pretrained/controlnet.bin \
+  --unet_path ControlNeXt-SVD-v2/pretrained/unet.bin \
+  --pose_video_path generated_pose_video/GOOD_MORNING_AFTERNOON_25073_36863_01435_20250804_160832.mp4 \
+  --ref_image_path ref_img_dict/03.jpeg \
+  --precision fp16 \
+  --decode_chunk_size 2 \
+  --enable_geglu_optimization \
+  --enable_xformers 
+```
+

text2pose.py

@@ -88,8 +88,10 @@ class SignLanguageQA:
             return None
             
         try:
-            data = np.load(npz_path, allow_pickle=True)
-            return data
+            npz_file = np.load(npz_path, allow_pickle=True)
+            npz_dict = {key: npz_file[key] for key in npz_file.files}
+            npz_file.close()
+            return npz_dict
         except Exception as e:
             print(f"   ⚠️  Failed to load NPZ: {e}")
             return None
@@ -411,26 +413,198 @@ class SignLanguageQA:
     def apply_y_offset(self, frame_data, y_offset):
         """Apply vertical offset to all pose points"""
         offset_data = copy.deepcopy(frame_data)
-        
+
         # Apply offset to bodies
         if 'bodies' in offset_data and offset_data['bodies'] is not None:
             offset_data['bodies'][:, 1] += y_offset
-        
+
         # Apply offset to hands
         if 'hands' in offset_data and offset_data['hands'] is not None:
             if len(offset_data['hands'].shape) == 3:
                 offset_data['hands'][:, :, 1] += y_offset
             else:
                 offset_data['hands'][:, 1] += y_offset
-        
+
         # Apply offset to faces
         if 'faces' in offset_data and offset_data['faces'] is not None:
             if len(offset_data['faces'].shape) == 3:
                 offset_data['faces'][:, :, 1] += y_offset
             else:
                 offset_data['faces'][:, 1] += y_offset
-        
+
         return offset_data
+
+    def normalize_pose_data(self, npz_data, reference_point='neck', scale_by='shoulders',
+                             target_shoulder_width=0.35, target_shoulder_y=0.45):
+        """
+        Normalize pose data to a standard coordinate system
+        This ensures different videos are in the same coordinate space
+
+        Args:
+            npz_data: NPZ data containing pose information
+            reference_point: 'neck' or 'hips' - point to center on (default: 'neck')
+            scale_by: 'shoulders' or 'height' - measure for scaling (default: 'shoulders')
+            target_shoulder_width: Desired horizontal distance between shoulders after normalization
+            target_shoulder_y: Desired Y coordinate for the shoulder line
+
+        Returns:
+            Normalized NPZ data (modifies in-place)
+        """
+        total_frames = self.get_total_frames(npz_data)
+        if total_frames == 0:
+            return npz_data
+
+        # Collect reference points, scale factors, and shoulder heights across all frames
+        reference_positions = []
+        scale_factors = []
+        shoulder_y_positions = []
+
+        for frame_num in range(1, total_frames + 1):
+            frame_data = self.get_frame_data(npz_data, frame_num)
+            if not frame_data or 'bodies' not in frame_data:
+                continue
+
+            bodies = frame_data['bodies']
+            if bodies is None or len(bodies) == 0:
+                continue
+
+            # Calculate reference point
+            if reference_point == 'neck':
+                # Neck is typically index 1 in body keypoints
+                if len(bodies) > 1 and bodies[1, 0] > 0 and bodies[1, 1] > 0:
+                    ref_pos = bodies[1, :2].copy()
+                else:
+                    continue
+            elif reference_point == 'hips':
+                # Hip center: average of left hip (11) and right hip (12)
+                if len(bodies) > 12:
+                    left_hip = bodies[11, :2]
+                    right_hip = bodies[12, :2]
+                    if left_hip[0] > 0 and right_hip[0] > 0:
+                        ref_pos = (left_hip + right_hip) / 2
+                    else:
+                        continue
+                else:
+                    continue
+            else:
+                continue
+
+            reference_positions.append(ref_pos)
+
+            # Calculate scale factor
+            if scale_by == 'shoulders':
+                # Shoulder width: distance between left shoulder (5) and right shoulder (2)
+                if len(bodies) > 5:
+                    left_shoulder = bodies[5, :2]
+                    right_shoulder = bodies[2, :2]
+                    if left_shoulder[0] > 0 and right_shoulder[0] > 0:
+                        shoulder_width = np.linalg.norm(left_shoulder - right_shoulder)
+                        if shoulder_width > 0.01:  # Avoid division by very small numbers
+                            scale_factors.append(shoulder_width)
+                            shoulder_mid_y = (left_shoulder[1] + right_shoulder[1]) / 2
+                            shoulder_y_positions.append(shoulder_mid_y)
+            elif scale_by == 'height':
+                # Approximate height: distance from neck to hip center
+                if len(bodies) > 12:
+                    neck = bodies[1, :2]
+                    left_hip = bodies[11, :2]
+                    right_hip = bodies[12, :2]
+                    if neck[1] > 0 and left_hip[0] > 0 and right_hip[0] > 0:
+                        hip_center = (left_hip + right_hip) / 2
+                        height = np.linalg.norm(neck - hip_center)
+                        if height > 0.01:
+                            scale_factors.append(height)
+
+        if not reference_positions or not scale_factors:
+            print("   ⚠️  Normalization skipped: insufficient valid frames")
+            return npz_data
+
+        # Use median reference position and scale factor for robustness
+        median_ref_pos = np.median(reference_positions, axis=0)
+        median_scale = np.median(scale_factors)
+
+        median_shoulder_y = np.median(shoulder_y_positions) if shoulder_y_positions else median_ref_pos[1]
+
+        print(f"   📏 Normalization: ref={reference_point}, scale={scale_by}")
+        print(f"      Reference position: ({median_ref_pos[0]:.3f}, {median_ref_pos[1]:.3f})")
+        print(f"      Median shoulder y: {median_shoulder_y:.3f}")
+        print(f"      Median shoulder width: {median_scale:.3f} → target {target_shoulder_width:.3f}")
+
+        # Determine uniform scaling factor so shoulders match the desired width
+        width_scale = target_shoulder_width / median_scale if median_scale > 0 else 1.0
+
+        # Neck-centered offset already moves the neck to 0.5; compute extra offset so shoulders sit at desired y
+        shoulder_delta_from_neck = median_shoulder_y - median_ref_pos[1]
+        normalized_shoulder_y = 0.5 + shoulder_delta_from_neck * width_scale
+        vertical_offset = target_shoulder_y - normalized_shoulder_y
+
+        # Apply normalization to all frames
+        for frame_num in range(1, total_frames + 1):
+            frame_key = f"frame_{frame_num:08d}"
+
+            # Normalize bodies
+            bodies_key = f"{frame_key}_bodies"
+            if bodies_key in npz_data:
+                bodies = npz_data[bodies_key].copy()
+                if bodies is not None and len(bodies) > 0:
+                    # Center on reference point
+                    for i in range(len(bodies)):
+                        if bodies[i, 0] > 0 and bodies[i, 1] > 0:
+                            bodies[i, :2] = (bodies[i, :2] - median_ref_pos) * width_scale + 0.5
+                            bodies[i, 1] += vertical_offset
+
+                    # Clamp to drawing range
+                    bodies[:, 0] = np.clip(bodies[:, 0], 0.0, 1.0)
+                    bodies[:, 1] = np.clip(bodies[:, 1], 0.0, 1.0)
+                    npz_data[bodies_key] = bodies
+
+            # Normalize hands
+            hands_key = f"{frame_key}_hands"
+            if hands_key in npz_data:
+                hands = npz_data[hands_key].copy()
+                if hands is not None:
+                    if len(hands.shape) == 3:
+                        # Multiple hands
+                        for h in range(hands.shape[0]):
+                            for i in range(hands.shape[1]):
+                                if hands[h, i, 0] > 0 and hands[h, i, 1] > 0:
+                                    hands[h, i, :2] = (hands[h, i, :2] - median_ref_pos) * width_scale + 0.5
+                                    hands[h, i, 1] += vertical_offset
+                    else:
+                        # Single array
+                        for i in range(len(hands)):
+                            if hands[i, 0] > 0 and hands[i, 1] > 0:
+                                hands[i, :2] = (hands[i, :2] - median_ref_pos) * width_scale + 0.5
+                                hands[i, 1] += vertical_offset
+
+                    hands[..., 0] = np.clip(hands[..., 0], 0.0, 1.0)
+                    hands[..., 1] = np.clip(hands[..., 1], 0.0, 1.0)
+                    npz_data[hands_key] = hands
+
+            # Normalize faces
+            faces_key = f"{frame_key}_faces"
+            if faces_key in npz_data:
+                faces = npz_data[faces_key].copy()
+                if faces is not None:
+                    if len(faces.shape) == 3:
+                        # Multiple faces
+                        for f in range(faces.shape[0]):
+                            for i in range(faces.shape[1]):
+                                if faces[f, i, 0] > 0 and faces[f, i, 1] > 0:
+                                    faces[f, i, :2] = (faces[f, i, :2] - median_ref_pos) * width_scale + 0.5
+                                    faces[f, i, 1] += vertical_offset
+                    else:
+                        # Single array
+                        for i in range(len(faces)):
+                            if faces[i, 0] > 0 and faces[i, 1] > 0:
+                                faces[i, :2] = (faces[i, :2] - median_ref_pos) * width_scale + 0.5
+                                faces[i, 1] += vertical_offset
+
+                    faces[..., 0] = np.clip(faces[..., 0], 0.0, 1.0)
+                    faces[..., 1] = np.clip(faces[..., 1], 0.0, 1.0)
+                    npz_data[faces_key] = faces
+
+        return npz_data
     
     def get_reference_head_top(self, ref_image_path, width, height):
         """Get the top of head position from reference image"""
@@ -560,11 +734,13 @@ class SignLanguageQA:
 
     def generate_video_from_glosses(self, gloss_list, output_path, fps=30, width=480, height=480,
                                     smoothing_frames=5, smoothing_method='none', npz_interpolation_frames=0,
-                                    scale_x=1.0, scale_y=1.0, hide_torso_lines=False, ref_image_path=None, draw_style='controlnext'):
+                                    scale_x=1.0, scale_y=1.0, hide_torso_lines=False, ref_image_path=None, draw_style='controlnext',
+                                    normalize_pose=True):
         """Generate video from gloss list with optional progressive scaling and head alignment
 
         Args:
             draw_style: 'controlnext' (default) or 'openpose' - drawing style for pose visualization
+            normalize_pose: Whether to normalize pose data to standard coordinate system (default: True)
         """
         # Create output directory
         output_path.parent.mkdir(parents=True, exist_ok=True)
@@ -642,6 +818,10 @@ class SignLanguageQA:
             print(f"   🎯 Centrality score: {centrality_score:.2f} (selected from {len(candidate_videos)} candidates)")
             print(f"   📊 Using frames {start_frame} to {end_frame} (out of {total_frames})")
 
+            # Apply normalization if enabled
+            if normalize_pose:
+                npz_data = self.normalize_pose_data(npz_data, reference_point='neck', scale_by='shoulders')
+
             # Store NPZ data for potential interpolation
             all_npz_data.append(npz_data)
 
@@ -1057,6 +1237,8 @@ def main():
                         help='Reference image path for head alignment')
     parser.add_argument('--draw-style', type=str, choices=['controlnext', 'openpose'], default='controlnext',
                         help='Drawing style for pose visualization. controlnext: advanced features with confidence-based coloring. openpose: simpler classic style. (default: controlnext)')
+    parser.add_argument('--normalize-pose', type=str, choices=['true', 'false'], default='true',
+                        help='Normalize pose data to standard coordinate system. This ensures different videos are in the same coordinate space, reducing inconsistency from different video sources. (default: true)')
 
     args = parser.parse_args()
 
@@ -1091,11 +1273,16 @@ def main():
         if args.draw_style != 'controlnext':
             print(f"🎨 Using {args.draw_style} drawing style")
 
+        if args.normalize_pose == 'true':
+            print(f"📏 Pose normalization enabled")
+        else:
+            print(f"⚠️  Pose normalization disabled")
+
         success = qa_system.generate_video_from_glosses(
             gloss_list, output_file, args.fps, args.width, args.height,
             args.smoothing_frames, args.smoothing, args.npz_interpolation,
             args.scale_x, args.scale_y, args.hide_torso_lines == 'true',
-            args.ref_image_path, args.draw_style
+            args.ref_image_path, args.draw_style, args.normalize_pose == 'true'
         )
 
         if not success:
@@ -1106,4 +1293,4 @@ def main():
 
 
 if __name__ == "__main__":
-    main()
+    main()