Upload 22 files

.gitattributes

@@ -38,3 +38,23 @@ ax630c/raft_steoro256x640_r1_npu2.axmodel filter=lfs diff=lfs merge=lfs -text
 ax630c/raft_steoro256x640_r4_npu2.axmodel filter=lfs diff=lfs merge=lfs -text
 ax650/raft_steoro256x640_r1.axmodel filter=lfs diff=lfs merge=lfs -text
 ax650/raft_steoro384x1280_r4.axmodel filter=lfs diff=lfs merge=lfs -text
+examples/left/000051_11.png filter=lfs diff=lfs merge=lfs -text
+examples/left/000058_11.png filter=lfs diff=lfs merge=lfs -text
+examples/left/000059_10.png filter=lfs diff=lfs merge=lfs -text
+examples/left/000121_10.png filter=lfs diff=lfs merge=lfs -text
+examples/left/000164_10.png filter=lfs diff=lfs merge=lfs -text
+examples/left/000167_11.png filter=lfs diff=lfs merge=lfs -text
+examples/left/000172_11.png filter=lfs diff=lfs merge=lfs -text
+examples/left/000179_10.png filter=lfs diff=lfs merge=lfs -text
+examples/left/000193_10.png filter=lfs diff=lfs merge=lfs -text
+examples/left/000195_10.png filter=lfs diff=lfs merge=lfs -text
+examples/right/000051_11.png filter=lfs diff=lfs merge=lfs -text
+examples/right/000058_11.png filter=lfs diff=lfs merge=lfs -text
+examples/right/000059_10.png filter=lfs diff=lfs merge=lfs -text
+examples/right/000121_10.png filter=lfs diff=lfs merge=lfs -text
+examples/right/000164_10.png filter=lfs diff=lfs merge=lfs -text
+examples/right/000167_11.png filter=lfs diff=lfs merge=lfs -text
+examples/right/000172_11.png filter=lfs diff=lfs merge=lfs -text
+examples/right/000179_10.png filter=lfs diff=lfs merge=lfs -text
+examples/right/000193_10.png filter=lfs diff=lfs merge=lfs -text
+examples/right/000195_10.png filter=lfs diff=lfs merge=lfs -text

infer.py

@@ -0,0 +1,125 @@
+import argparse
+import cv2
+# from PIL import Image
+enable_cv2 = True
+
+import numpy as np
+import matplotlib.pyplot as plt
+from axengine import InferenceSession
+
+def bilinear_resize_numpy(array, new_h, new_w):
+    h, w = array.shape
+    x_ratio = w / new_w
+    y_ratio = h / new_h
+    
+    resized = np.zeros((new_h, new_w), dtype=array.dtype)
+    
+    for i in range(new_h):
+        for j in range(new_w):
+            x = j * x_ratio
+            y = i * y_ratio
+            
+            x_floor = int(x)
+            y_floor = int(y)
+            x_ceil = min(x_floor + 1, w - 1)
+            y_ceil = min(y_floor + 1, h - 1)
+            
+            dx = x - x_floor
+            dy = y - y_floor
+            
+            a = array[y_floor, x_floor]
+            b = array[y_floor, x_ceil]
+            c = array[y_ceil, x_floor]
+            d = array[y_ceil, x_ceil]
+            
+            resized[i, j] = a * (1 - dx) * (1 - dy) + b * dx * (1 - dy) + c * (1 - dx) * dy + d * dx * dy
+    
+    return resized
+
+
+def resize_disp(disp, target_width, target_height, use_cv2=True):
+    if use_cv2:
+        disp = cv2.resize(disp, (target_width, target_height))
+    else:
+        # This implementation is slower than cv2.resize
+        disp = bilinear_resize_numpy(disp, target_height, target_width)
+
+    return disp
+
+def load_and_preprocess_image(image_path, target_width, target_height, use_cv2=True):
+
+    if use_cv2:
+        img = cv2.imread(image_path)
+        img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB) 
+        orig_height, orig_width = img.shape[:2]
+        img_resized = cv2.resize(img, (target_width,target_height) )
+        img_batch = img_resized[None]
+    else:
+        img = Image.open(image_path).convert('RGB')
+        orig_width, orig_height = img.size
+        img_resized = img.resize((target_width, target_height))
+        img_array = np.array(img_resized)  #
+        img_batch = img_array[None]        # 
+    
+    return img_batch, (orig_height, orig_width)
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--left",
+        type=str,
+        required=True,
+        help="Path to left image.",
+    )
+    parser.add_argument(
+        "--right",
+        type=str,
+        required=True,
+        help="Path to right image.",
+    )
+    parser.add_argument(
+        "--model",
+        type=str,
+        required=True,
+        help="Path to ONNX model.",
+    )
+    parser.add_argument(
+        "--width",
+        type=int,
+        required=True,
+        help="Width of input image.",
+    )
+    parser.add_argument(
+        "--height",
+        type=int,
+        required=True,
+        help="Height of input image.",
+    )
+    
+    return parser.parse_args()
+
+
+def infer(left: str, right: str, model: str, width:int, height:int):
+    
+    image_left, (orig_h_left, orig_w_left) = load_and_preprocess_image(left, width, height, use_cv2=enable_cv2)
+    image_right, (orig_h_right, orig_w_right) = load_and_preprocess_image(right, width, height, use_cv2=enable_cv2)
+
+    assert orig_h_left == orig_h_right and orig_w_left == orig_w_right
+
+    session = InferenceSession.load_from_model(model)
+    flow_up = session.run(input_feed={"x1":image_left, "x2":image_right})["output"]
+
+    # Please use opencv whenever possible
+    flow_up = resize_disp(flow_up[0,0], orig_w_left, orig_h_left, use_cv2=enable_cv2)
+
+    flow_up *= orig_w_left/width
+    output = np.abs(flow_up)
+    
+    plt.imsave(f"output-ax.png", output, cmap='jet')
+
+    return output
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    infer(**vars(args))

infer_onnx.py

@@ -0,0 +1,100 @@
+import argparse
+import cv2
+import numpy as np
+import onnxruntime as ort
+import matplotlib.pyplot as plt
+
+def parse_args() -> argparse.Namespace:
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--left",
+        type=str,
+        required=True,
+        help="Path to left image.",
+    )
+    parser.add_argument(
+        "--right",
+        type=str,
+        required=True,
+        help="Path to right image.",
+    )
+    parser.add_argument(
+        "--model",
+        type=str,
+        required=True,
+        help="Path to ONNX model.",
+    )
+    
+    return parser.parse_args()
+
+
+def infer(left: str, right: str, model: str):
+    # 自动检测可用的执行提供者，优先使用 CUDA，否则使用 CPU
+    available_providers = ort.get_available_providers()
+    if "CUDAExecutionProvider" in available_providers:
+        providers = ["CUDAExecutionProvider", "CPUExecutionProvider"]
+        print("使用 CUDA 执行提供者")
+    else:
+        providers = ["CPUExecutionProvider"]
+        print("使用 CPU 执行提供者 (CUDA 不可用)")
+
+    print(f"正在加载模型: {model}")
+    session = ort.InferenceSession(
+        model, providers=providers
+    )
+    print("模型加载完成")
+    
+    input_info = []
+    input_tensors = session.get_inputs()
+    for tensor in input_tensors:
+        info = dict(name=tensor.name , type=tensor.type, shape=tensor.shape)
+        input_info.append(info)
+
+    H,W = input_info[0]['shape'][2:4]
+    print(f"输入尺寸: {H}x{W}")
+
+    print(f"正在读取左图: {left}")
+    left_raw = cv2.imread(left)
+    if left_raw is None:
+        raise ValueError(f"无法读取左图: {left}")
+    image_left = cv2.cvtColor(left_raw, cv2.COLOR_BGR2RGB) 
+    orig_h_left, orig_w_left = image_left.shape[:2]
+    print(f"左图原始尺寸: {orig_h_left}x{orig_w_left}")
+    image_left = cv2.resize(image_left, (W,H) )
+   
+    image_left = image_left.transpose(2,0,1)
+    image_left = image_left[None].astype(np.float32)
+
+    print(f"正在读取右图: {right}")
+    right_raw = cv2.imread(right)
+    if right_raw is None:
+        raise ValueError(f"无法读取右图: {right}")
+    image_right = cv2.cvtColor(right_raw, cv2.COLOR_BGR2RGB) 
+    orig_h_right, orig_w_right = image_right.shape[:2]
+    print(f"右图原始尺寸: {orig_h_right}x{orig_w_right}")
+    image_right = cv2.resize(image_right, (W,H) )
+   
+    image_right = image_right.transpose(2,0,1)
+    image_right = image_right[None].astype(np.float32)
+
+    assert orig_h_left == orig_h_right and orig_w_left == orig_w_right
+
+    print("正在进行推理...")
+    flow_up = session.run(None, {input_info[0]['name']: image_left, input_info[1]['name']:image_right})[0]
+    print("推理完成")
+    
+    flow_up = cv2.resize(flow_up[0,0], (orig_w_left, orig_h_left))
+    flow_up *= orig_w_left/W
+    
+    output = np.abs(flow_up)
+    
+    print(f"正在保存结果到 output-onnx.png")
+    plt.imsave(f"output-onnx.png", output, cmap='jet')
+    print("推理完成，结果已保存到 output-onnx.png")
+
+    return output
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    infer(**vars(args))