xttmp 2.3.0.1__tar.gz → 2.3.0.3__tar.gz

{xttmp-2.3.0.1/src/xttmp.egg-info → xttmp-2.3.0.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: xttmp
-Version: 2.3.0.1
+Version: 2.3.0.3
 Summary: eXtremely Tiny Target - Motion Perception
 Author-email: Shawn MX <mingshuoxu@hotmail.com>
 Project-URL: Homepage, https://github.com/MingshuoXu/Small-Target-Motion-Detectors
@@ -16,15 +16,16 @@ Classifier: Programming Language :: Python :: 3.9
 Classifier: Programming Language :: Python :: 3.10
 Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
-Classifier: Topic :: Scientific/Engineering :: Image Recognition
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Requires-Python: >=3.8
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: matplotlib
 Requires-Dist: opencv-python
 Requires-Dist: scipy
-Requires-Dist: torch>=2.5.0
-Requires-Dist: torchvision>=0.20.0
+Provides-Extra: torch
+Requires-Dist: torch>=2.5.0; extra == "torch"
+Requires-Dist: torchvision>=0.20.0; extra == "torch"
 Dynamic: license-file
 
 # Small Target Motion Detectors, Version 2.3 (XTT-MP: Extremely Tiny Target - Motion Perception)
@@ -66,8 +67,18 @@ Built with modularity and extensibility in mind, XTT-MP provides a robust suite
 - After `pip install xttmp`, use the installed code and bring your own input data, or run from a repository checkout to access the bundled examples.
 
 ### Via PyPI
+#### CPU
+```bash
+pip install xttmp[torch]
+```
+
+#### NVIDIA GPU (CUDA 12.6)
+```bash
+pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126
+```
+
+### Running the GUI Demo
 ```bash
-pip install xttmp
 xttmp_gui
 ```
 

{xttmp-2.3.0.1 → xttmp-2.3.0.3}/README.md

@@ -37,8 +37,18 @@ Built with modularity and extensibility in mind, XTT-MP provides a robust suite
 - After `pip install xttmp`, use the installed code and bring your own input data, or run from a repository checkout to access the bundled examples.
 
 ### Via PyPI
+#### CPU
+```bash
+pip install xttmp[torch]
+```
+
+#### NVIDIA GPU (CUDA 12.6)
+```bash
+pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126
+```
+
+### Running the GUI Demo
 ```bash
-pip install xttmp
 xttmp_gui
 ```
 

{xttmp-2.3.0.1 → xttmp-2.3.0.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "xttmp"
-version = "2.3.0.1"
+version = "2.3.0.3"
 description = "eXtremely Tiny Target - Motion Perception"
 readme = "README.md"
 requires-python = ">=3.8"
@@ -20,14 +20,12 @@ classifiers = [
     "Programming Language :: Python :: 3.10",
     "Programming Language :: Python :: 3.11",
     "Programming Language :: Python :: 3.12",
-    "Topic :: Scientific/Engineering :: Image Recognition",
+    "Topic :: Scientific/Engineering :: Artificial Intelligence",
 ]
 dependencies = [
     "matplotlib",
     "opencv-python",
     "scipy",
-    "torch>=2.5.0",
-    "torchvision>=0.20.0",
 ]
 
 [project.urls]
@@ -47,6 +45,12 @@ where = ["src"]
 [tool.setuptools.package-data]
 "xttmp" = ["util/*.ico"]
 
+[project.optional-dependencies]
+torch = [
+    "torch>=2.5.0",
+    "torchvision>=0.20.0"
+]
+
 [dependency-groups]
 dev = []
 

{xttmp-2.3.0.1 → xttmp-2.3.0.3}/src/xttmp/demo/inference_gui.py

@@ -12,19 +12,14 @@ file_path = os.path.realpath(__file__)
 py_pkg_path = os.path.dirname(os.path.dirname(os.path.dirname(file_path)))
 sys.path.append(py_pkg_path)
 
-try:
-    from xttmp.util.iostream import ( # type: ignore
-                    ModelAndInputSelectorGUI,
-                    FrameIterator,
-                    FrameVisualizer,
-                )
-    from xttmp.util.compute_module import PostProcessing # type: ignore
-    from xttmp.api import ( # type: ignore
-        instancing_model,
-    ) 
-except ImportError as e:
-    raise ImportError("Failed to import required modules. "
-                      "Ensure that the 'xttmp' package is correctly installed.") from e
+
+from xttmp.util.iostream import ( # type: ignore
+                XTTMP_GUI,
+                FrameIterator,
+                FrameVisualizer,
+            )
+from xttmp.api import instancing_model
+
 
 # configure logging
 logging.basicConfig(level=logging.INFO,
@@ -32,18 +27,15 @@ logging.basicConfig(level=logging.INFO,
 logger = logging.getLogger(__name__)
 
 class StmdGui:
-    def __init__(self, device='cpu', show_threshold: float = 0.8, get_top_num: int = 1):
+    def __init__(self):
         """ Initialize STMD GUI """
-        self.device = device
-        self.show_threshold = show_threshold
-        self.get_top_num = get_top_num
-        self.ModelAndInputSelectorGUI = ModelAndInputSelectorGUI
+        self.device = None
+        self.ModelAndInputSelectorGUI = XTTMP_GUI
         self.FrameIterator = FrameIterator
         self.FrameVisualizer = FrameVisualizer
-        self.PostProcessing = PostProcessing
+        self.post_processor = None
         self.instancing_model = instancing_model
 
-
     def _get_user_input(self) -> tuple:
         """ get user input """
         root = tk.Tk()
@@ -93,21 +85,17 @@ class StmdGui:
                 logger.info("User cancelled input.")
                 return 
 
-            model_name, opt1, opt2, is_stepping = user_input
+            model_name, opt1, opt2, is_stepping, device, post_processor, show_threshold = user_input
+            self.post_processor = post_processor
+            self.device = device
             reader = self._create_frame_reader(opt1, opt2)
-            model = self.instancing_model(model_name, device=self.device)
-            post_processor = self.PostProcessing(
-                device=self.device,
-                nms_radio=8,
-                get_top_num=self.get_top_num,
-            )
+            model = self.instancing_model(model_name, device)
 
             visualizer = self.FrameVisualizer(
                 window_name=model_name,
-                result_index_type="dots",
                 win_width=reader.img_width,
                 win_height=reader.img_height,
-                conf_threshold=self.show_threshold,
+                conf_threshold=show_threshold,
             )
             if is_stepping:
                 visualizer.paused = True
@@ -117,16 +105,17 @@ class StmdGui:
                 if not is_valid:
                     break
 
-                if self.device != 'cpu' and torch.cuda.is_available():
+                if self.device == 'cuda':
                     torch.cuda.synchronize()
                 time_start = time.perf_counter()
                 result = model(gray_tensor)
-                if self.device != 'cpu' and torch.cuda.is_available():
+                if self.device == 'cuda':
                     torch.cuda.synchronize()
                 run_time = time.perf_counter() - time_start
 
-                dots = post_processor(result['response'], result.get('direction'))
-                if not visualizer.update(color_img, result=dots, process_time=run_time):
+                post_res = post_processor(result['response'], result.get('direction'))
+                show_str = f'{self.device} : {run_time*1000:.1f} ms'
+                if not visualizer.update(color_img, result=post_res, show_str=show_str):
                     break
 
         except Exception as e:
@@ -139,10 +128,7 @@ class StmdGui:
                 reader.release()
             logger.info("Shutdown completed")
 
-def main(show_threshold: float = 0, get_top_num: int = 10):
-    DEVICE = 'cuda' if torch.cuda.is_available() else 'cpu'
-    obj = StmdGui(DEVICE, show_threshold = show_threshold, get_top_num = get_top_num)
-    obj.run()
 
 if __name__ == "__main__":
-    main(get_top_num = 20)
+    obj = StmdGui()
+    obj.run()

{xttmp-2.3.0.1 → xttmp-2.3.0.3}/src/xttmp/main.py

@@ -2,6 +2,13 @@ from pathlib import Path
 import subprocess
 import sys
 
+try:
+    import torch
+except ImportError:
+    raise ImportError(
+        "Please install PyTorch first. "
+        "See https://pytorch.org/get-started/locally/"
+    )
 
 def main():
 	script_path = Path(__file__).resolve().parent / 'demo' / 'inference_gui.py'

{xttmp-2.3.0.1 → xttmp-2.3.0.3}/src/xttmp/util/compute_module.py

@@ -6,67 +6,6 @@ import torch
 import torch.nn.functional as F
 
 
-def compute_temporal_conv(iptCell, kernel, pointer=None):
-    """
-    Computes temporal convolution.
-
-    Parameters:
-    - iptCell: A list of arrays where each element has the same dimension.
-    - kernel: A vector representing the convolution kernel.
-    - headPointer: Head pointer of the input cell array (optional).
-
-    Returns:
-    - optMatrix: The result of the temporal convolution.
-    """
-
-    # Default value for headPointer
-    if pointer is None:
-        pointer = len(iptCell) - 1
-
-    # Initialize output matrix
-    if iptCell[pointer] is None:
-        return None
-
-    # Ensure kernel is a vector
-    kernel = np.squeeze(kernel)
-    if not np.ndim(kernel) == 1:
-        raise ValueError('The kernel must be a vector.')
-
-    # Determine the lengths of input cell array and kernel
-    k1 = len(iptCell)
-    k2 = len(kernel)
-    length = min(k1, k2)
-
-    if isinstance(iptCell[pointer], np.ndarray):
-        optMatrix = np.zeros_like(iptCell[pointer])
-    elif isinstance(iptCell[pointer], torch.Tensor):
-        optMatrix = torch.zeros_like(iptCell[pointer])
-    # Perform temporal convolution
-    for t in range(length):
-        j = (pointer - t) % k1
-        if abs(kernel[t]) > 1e-16 and iptCell[j] is not None:
-            optMatrix += iptCell[j] * kernel[t]
-
-    return optMatrix
-
-
-def compute_circularlist_conv(circularCell, temporalKernel):
-    """
-    Compute the convolution of a circular cell with a temporal kernel.
-    
-    Args:
-    - circularCell: The circular cell data.
-    - temporalKernel: The temporal kernel data.
-    
-    Returns:
-    - opt_matrix: The result of the convolution.
-    """
-    optMatrix = compute_temporal_conv(circularCell, 
-                                      temporalKernel, 
-                                      circularCell.pointer )
-    return optMatrix
-
-
 def compute_response(ipt):
     """
     Computes the maximum response from multiple inputs.
@@ -233,139 +172,17 @@ class AreaNMS:
         return matrix * (matrix == local_max)
     
 
-def get_top_k_torch(response_tensor, direction_tensor, k=1000):
-    """
-    输入: 
-        response_tensor: (..., H, W) 任意维度的 Tensor
-        direction_tensor: (..., H, W) 形状需与 response 匹配 (可选)
-    输出: 
-        torch.Tensor: shape=(M, 4), dtype=float32, 其中 M <= k
-        格式: [[x, y, response, direction], ...]
-    """
-    # 1. 获取维度
-    H, W = response_tensor.shape[-2:]
-    k = min(k, H * W)
-
-    # 2. 展平 (Flatten)
-    # view(-1) 零拷贝，极快
-    flat_response = response_tensor.view(-1)
-
-    # 3. TopK (GPU 上极速排序)
-    top_vals, top_indices = torch.topk(flat_response, k=k)
-
-    # 4. 过滤掉 <= 0 的值 ---
-    # 创建掩码：只保留大于 0 的值
-    mask = top_vals > 0
-    
-    # 如果全都是 0，直接返回空数组，避免后续报错
-    if not mask.any():
-        return torch.empty((0, 4))
-
-    # 应用掩码，缩减 tensor 长度
-    top_vals = top_vals[mask]
-    top_indices = top_indices[mask]
-    # ------------------------------------
-
-    # 5. 计算坐标 (x, y)
-    # 此时计算量已经减少，只计算非零点
-    top_y = top_indices.div(W, rounding_mode='floor').float() 
-    top_x = (top_indices % W).float()                        
-
-    # 6. 获取 Direction
-    if direction_tensor is not None and direction_tensor.numel() > 0:
-        flat_direction = direction_tensor.view(-1)
-        # 注意：这里使用过滤后的 top_indices
-        top_dirs = flat_direction[top_indices]
-    else:
-        top_dirs = torch.empty_like(top_vals).fill_(float('nan'))
-
-    # 7. 堆叠 (Stack) -> (M, 4)
-    result_tensor = torch.stack([top_x, top_y, top_vals, top_dirs], dim=1)
-
-    return result_tensor
-
-
-def get_top_k_numpy(response_array, direction_array=None, k=1000):
-    """
-    输入: 
-        response_array: (..., H, W) numpy.ndarray
-        direction_array: (..., H, W) (可选)
-    输出: 
-        numpy.ndarray: shape=(M, 4), dtype=float32, 其中 M <= k
-        格式: [[x, y, response, direction], ...]
-    """
-    # 1. 获取维度
-    shape = response_array.shape
-    H, W = shape[-2:]
-    
-    # 零拷贝展平
-    flat_response = response_array.ravel()
-    k = min(k, flat_response.size)
-
-    # 2. TopK 核心优化 (O(N))
-    # argpartition 找出最大的 k 个 (无序)
-    unsorted_top_indices = np.argpartition(flat_response, -k)[-k:]
-    unsorted_top_vals = flat_response[unsorted_top_indices]
-    
-    # 3. 局部排序 (O(k log k))
-    # argsort 默认升序，[::-1] 翻转为降序
-    sort_idx = np.argsort(unsorted_top_vals)[::-1]
-    
-    # 获取排序后的 Top K 索引和值
-    top_indices = unsorted_top_indices[sort_idx]
-    top_vals = unsorted_top_vals[sort_idx]
-
-    # --- [关键修改] 4. 过滤掉 <= 0 的值 ---
-    # 创建掩码
-    mask = top_vals > 0
-    
-    # 极速判断：如果没有有效值，直接返回空数组
-    # np.any() 很快
-    if not np.any(mask):
-        return np.empty((0, 4), dtype=np.float32)
-        
-    # 应用掩码 (切片操作，只保留有效值)
-    # 因为 k 通常不大 (比如 1000)，这里的拷贝开销可忽略不计
-    top_vals = top_vals[mask]
-    top_indices = top_indices[mask]
-    
-    # 更新实际数量 M
-    M = top_vals.size
-    # ------------------------------------
-
-    # 5. 计算坐标 (x, y)
-    # 只对过滤后的索引计算，节省算力
-    top_y, top_x = np.unravel_index(top_indices, (H, W))
-
-    # 6. 获取 Direction
-    if direction_array is not None and direction_array.size > 0:
-        flat_direction = direction_array.ravel()
-        top_dirs = flat_direction[top_indices]
-    else:
-        top_dirs = np.full(M, np.nan, dtype=np.float32)
-
-    # 7. 堆叠结果
-    # 分配恰好大小为 M 的内存
-    result = np.empty((M, 4), dtype=np.float32)
-    result[:, 0] = top_x       # x
-    result[:, 1] = top_y       # y
-    result[:, 2] = top_vals    # response
-    result[:, 3] = top_dirs    # direction
-
-    return result
-
-
 class PostProcessing:
     """
     Post-processing class to apply AreaNMS, get top K, and return list format.
     """
 
-    def __init__(self, device='cpu', nms_radio = 8, get_top_num=1000):
+    def __init__(self, nms_radio = 8, get_top_num=1000):
         """
         Args:
-            device (str): Computing device ('cpu' or 'cuda').
+            nms_radio (int): Radius for AreaNMS.
+            get_top_num (int): Number of top points to extract.
         """
-        self.device = device
         self.area_nms = AreaNMS(radio=nms_radio)
         self.get_top_num = get_top_num
 
@@ -389,7 +206,7 @@ class PostProcessing:
         """
         nms_response = self.area_nms(response)
 
-        res = get_top_k_torch(nms_response, 
+        res, _ = get_top_k_torch(nms_response, 
                                 direction, 
                                 k=self.get_top_num)
         if res.shape[0] == 0:
@@ -400,3 +217,144 @@ class PostProcessing:
                 res[:, 2] /= max_score
 
         return res
+
+
+@torch.no_grad()
+def gen_bboxes_around_points(results, box_size=16, shift_ratio=0.3):
+    """Generate initial bboxes around detected motion points.
+    
+    Args:
+        results: (N, 4) -> [x, y, response, direction]
+        box_size: int, box size
+        
+    Returns:
+        [[x1, y1, x2, y2], ... ] (N, 4) tensors
+    """
+    N = results.shape[0]
+
+    if N == 0:
+        return torch.empty((0, 4), device=results.device, dtype=torch.int)
+    
+    rear_x, rear_y, direction = results[:, 0], results[:, 1], results[:, 3]
+
+    radius = box_size * 0.5
+    shift_mag = box_size * shift_ratio
+
+    # 1. 计算偏移量 (dx, dy)，根据方向和预设的 shift_mag
+    dx = torch.cos(direction) * shift_mag
+    dy = -torch.sin(direction) * shift_mag
+
+    # 2. 一次性将 NaN 偏移量替换为 0.0
+    dx = torch.nan_to_num(dx, nan=0.0)
+    dy = torch.nan_to_num(dy, nan=0.0)
+
+    # 3. 计算中心点
+    center_x = rear_x + dx
+    center_y = rear_y + dy
+
+    x1 = (center_x - radius)
+    y1 = (center_y - radius)
+    x2 = (center_x + radius)
+    y2 = (center_y + radius)
+
+    # Stack as (N, 4) -> [x1, y1, x2, y2]
+    return torch.stack([x1, y1, x2, y2], dim=1)
+
+
+@torch.no_grad()
+def get_top_k_torch(response_tensor, direction_tensor=None, k=100):
+    """
+    Extract the top-k points with highest responses from feature maps, filtering out non-positive values.
+    
+    Args:
+        response_tensor (torch.Tensor): The response map tensor of shape (B, 1, H, W) or (B, H, W).
+        direction_tensor (torch.Tensor, optional): The corresponding direction map tensor of 
+            shape (B, 1, H, W) or (B, H, W). Must match response_tensor's shape. Defaults to None.
+        k (int, optional): The maximum number of top points to extract per batch. Defaults to 100.
+        
+    Returns:
+        Tuple[torch.Tensor, torch.Tensor]: 
+            - results (torch.Tensor): A tensor of shape (M, 4) containing the valid extracted points 
+            across the entire batch. M <= B * k. Each row is formatted as [x, y, response, direction].
+            - batch_ids (torch.Tensor): A 1D tensor of shape (M,) containing the corresponding 
+            batch index (from 0 to B-1) for each point in `results`. dtype is torch.long.
+    """
+    B, _, H, W = response_tensor.shape
+    k = min(k, H * W)
+    device = response_tensor.device
+
+    # 1. Flatten -> (B, H*W)
+    flat_response = response_tensor.reshape(B, -1)
+
+    # 2. TopK -> top_vals and top_indices are both (B, k)
+    top_vals, top_indices = torch.topk(flat_response, k=k, dim=-1)
+
+    # 3. Get Direction -> (B, k)
+    if direction_tensor is not None and direction_tensor.numel() > 0:
+        flat_direction = direction_tensor.reshape(B, -1)
+        top_dirs = torch.gather(flat_direction, dim=-1, index=top_indices)
+    else:
+        top_dirs = torch.full_like(top_vals, float('nan'))
+
+    # 4. Calculate coordinates (x, y) -> (B, k)
+    top_y = top_indices.div(W, rounding_mode='floor').float() 
+    top_x = (top_indices % W).float()                        
+
+    # 5. Stack -> merge on the last dimension, shape becomes (B, k, 4)
+    stacked = torch.stack([top_x, top_y, top_vals, top_dirs], dim=-1)
+
+    # 6. Generate Mask -> (B, k)
+    mask = top_vals > 0
+
+    # 7. Split and filter by Batch
+    result_list = []
+    batch_id_list = []
+    
+    for i in range(B):
+        batch_mask = mask[i] # Get the mask for the i-th batch
+        
+        # Apply mask: [k, 4] -> [M_i, 4]
+        valid_stacked = stacked[i][batch_mask] 
+        result_list.append(valid_stacked)
+        
+        # Create a batch index tensor of shape (M_i,) filled with the current batch index 'i'
+        batch_id_list.append(torch.full((valid_stacked.shape[0],), i, device=device, dtype=torch.long))
+    
+    # Concatenate all valid items into continuous tensors
+    return torch.cat(result_list, dim=0), torch.cat(batch_id_list, dim=0)
+
+
+@torch.no_grad()
+def get_STMD_region_proposal(response_tensor, direction_tensor=None, top_k=1, box_size=16, spatial_scale=1.0, shift_ratio=0.3):
+    nms_win = int(box_size * spatial_scale) | 1  # 确保是奇数
+    score_mask = F.max_pool2d(response_tensor, kernel_size=nms_win, stride=1, padding=nms_win//2)
+    
+    nms_response_tensor = torch.where(response_tensor == score_mask, response_tensor, 0.0)
+
+    vSTMD_res, batch_id = get_top_k_torch(nms_response_tensor, direction_tensor, k=top_k)
+
+    if spatial_scale > 1:
+        vSTMD_res[:, :2] *= spatial_scale   # 将坐标放大回原图尺度
+
+    bboxes = gen_bboxes_around_points(vSTMD_res, box_size, shift_ratio)
+
+    return vSTMD_res, bboxes, batch_id
+
+
+@torch.no_grad()
+def bbox_post_processing(top_k=1, box_size=16, spatial_scale=1.0, shift_ratio=0.3):
+
+    def post_process_func(
+        response_tensor,
+        direction_tensor=None
+    ):
+        vSTMD_res, bboxes, _ =  get_STMD_region_proposal(response_tensor,
+                                                        direction_tensor,
+                                                        top_k=top_k,
+                                                        box_size=box_size,
+                                                        spatial_scale=spatial_scale,
+                                                        shift_ratio=shift_ratio )
+        return torch.cat([bboxes, vSTMD_res[..., 2:3]], dim=1)
+    
+    return post_process_func
+

{xttmp-2.3.0.1 → xttmp-2.3.0.3}/src/xttmp/util/iostream.py

@@ -3,6 +3,7 @@ import re
 from pathlib import Path
 import logging
 from typing import Optional, List, Union, Tuple, Any
+from functools import partial
 
 import cv2
 import numpy as np
@@ -12,6 +13,7 @@ import torch
 
 
 from .. import model
+from .compute_module import PostProcessing, bbox_post_processing
 
 
 # Get the full path of this file
@@ -215,7 +217,6 @@ class FrameIterator:
 
 class FrameVisualizer:
     def __init__(self, window_name="Visualizer", 
-                 result_index_type="matrix",
                  win_width=None, win_height=None, 
                  is_headless=False,
                  conf_threshold=0.8 # 阈值参数
@@ -225,7 +226,6 @@ class FrameVisualizer:
         :param conf_threshold: 可视化过滤的相对阈值 (0.0 ~ 1.0)
         """
         self.window_name = window_name
-        self.result_index_type = result_index_type  # "matrix", "dots", "bbox"
         self.win_width = win_width or 800
         self.win_height = win_height or 600
         self.is_headless = is_headless
@@ -260,24 +260,26 @@ class FrameVisualizer:
         self.save_output = True
         print(f">>> Video writer initialized: {output_path}")
 
-    def update(self, frame, result=None, direction=None, annotation=None, process_time=None) -> bool:
+    def update(self, frame, result=None, direction=None, annotation=None, show_str=None) -> bool:
         if frame is None:
             return False
 
         # --- 绘制逻辑 ---
         # 即使 result 是空的，只要不为 None 也可以处理
         if result is not None:
-            if self.result_index_type == "matrix":
+            if result.dim() == 4:
                 self._draw_matrix(frame, result, direction, self.conf_threshold)
-            elif self.result_index_type == "dots": 
+            elif result.shape[1] == 4: 
                 result = result.cpu().numpy() if isinstance(result, torch.Tensor) else result
                 self._draw_dots(frame, result, self.conf_threshold)
-            elif self.result_index_type == "bbox": 
+            elif result.shape[1] == 5: 
                 self._draw_bbox(frame, result, self.conf_threshold, annotation)
-
+            device_str = f'{result.device}'
+        else:
+            device_str = 'Time'
         # --- 信息显示 ---
-        if process_time is not None:
-            cv2.putText(frame, f'Time: {process_time*1000:.1f} ms',
+        if show_str is not None and show_str != '':
+            cv2.putText(frame, str(show_str),
                         (10, 30), cv2.FONT_HERSHEY_SIMPLEX, 0.8,
                         (0, 255, 0), 2, cv2.LINE_AA)
             
@@ -344,10 +346,12 @@ class FrameVisualizer:
     @staticmethod
     def _draw_matrix(frame, matrix, direction_map, threshold):
         """处理 Matrix 格式 (Heatmap)"""
-        if np.max(matrix) <= 0: return
+        if torch.max(matrix) <= 0: return
 
         # np.where 返回 (rows, cols) 即 (y, x)
-        rows, cols = np.where(matrix > threshold)
+        _, _, rows, cols = torch.where(matrix > threshold)
+        rows = rows.cpu().numpy()
+        cols = cols.cpu().numpy()
         
         # 画点
         for r, c in zip(rows, cols):
@@ -358,7 +362,7 @@ class FrameVisualizer:
         # 画箭头
         if direction_map is not None and len(rows) > 0:
             # 确保 direction_map 维度匹配，这里假设是同样大小的矩阵
-            valid_dirs = direction_map[rows, cols]
+            valid_dirs = direction_map[0, 0, rows, cols]
             
             # 过滤 NaN
             valid_mask = ~np.isnan(valid_dirs)
@@ -450,9 +454,9 @@ class ModelSelectorGUI:
         self.modelLabel = ttk.Label(self.root, text="Select a model:", width = 15)
         self.modelLabel.grid(row=0, column=0, padx=10, pady=10)
 
-        self.modelCombobox = ttk.Combobox(self.root, values=modelList, width = 30)
+        self.modelCombobox = ttk.Combobox(self.root, values=modelList, width = 25)
         self.modelCombobox.current(11)
-        self.modelCombobox.grid(row=0, column=1, columnspan=2, padx=10, pady=10)
+        self.modelCombobox.grid(row=0, column=1, columnspan=2, pady=10, sticky='w')
         
 
 class InputSelectorGUI:
@@ -476,25 +480,24 @@ class InputSelectorGUI:
         self.endImgName = None
 
     def create_gui(self):
-        self.inputTypeLabel = ttk.Label(self.root, text="Select input from:", width = 15)
-        self.inputTypeLabel.grid(row=1, column=0, padx=10, pady=10)
+        self.inputTypeLabel = ttk.Label(self.root, text="Input Type:", width = 15)
+        self.inputTypeLabel.grid(row=1, column=0, padx=10, pady=10, sticky='w')
 
         self.selectedOption = tk.IntVar(value=0)
 
-
-        self.vidLabel = ttk.Radiobutton(self.root, 
-                                        text='Video stream', 
-                                        variable=self.selectedOption,
-                                        value=1, 
-                                        command=self.select_vidstream)
-        self.vidLabel.grid(row=1, column=2, padx=10, pady=10)
-        
         self.imgLabel = ttk.Radiobutton(self.root, 
-                                        text='Image stream', 
+                                        text='Image Sequence', 
                                         variable=self.selectedOption,
                                         value=2, 
                                         command=self.select_imgstream)
-        self.imgLabel.grid(row=1, column=1, padx=10, pady=10)
+        self.imgLabel.grid(row=1, column=2, padx=10, pady=10, sticky="w")
+        
+        self.vidLabel = ttk.Radiobutton(self.root, 
+                                        text='Video', 
+                                        variable=self.selectedOption,
+                                        value=1, 
+                                        command=self.select_vidstream)
+        self.vidLabel.grid(row=1, column=1, padx=10, pady=10, sticky="w")
 
     def select_vidstream(self):
         self.imgSelectFolder = None
@@ -503,16 +506,16 @@ class InputSelectorGUI:
         for element in self.imgElement.values():
             element.destroy()
 
-        self.vidElement['lblVidIndicate'] = ttk.Label(self.root, text= 'Video\'s path:',  width = 15)
-        self.vidElement['lblVidIndicate'].grid(row=2, column=0, padx=10, pady=30)
+        self.vidElement['lblVidIndicate'] = ttk.Label(self.root, text= 'Video\'s path:')
+        self.vidElement['lblVidIndicate'].grid(row=2, column=1, padx=10, pady=30, sticky='w')
         self.vidElement['lblVidPath'] = ttk.Label(self.root, 
                                            text="Waiting for the selection",
                                            wraplength=220
                                            )
-        self.vidElement['lblVidPath'].grid(row=2, column=1, columnspan=2, padx=10, pady=10)
+        self.vidElement['lblVidPath'].grid(row=2, column=2, padx=10, pady=10, sticky='w')
         
         self.vidElement['btn'] = ttk.Button(self.root, text="Select a video", command=self._clicked_vid)
-        self.vidElement['btn'].grid(row=3, column=2, padx=10, pady=10)
+        self.vidElement['btn'].grid(row=3, column=2, padx=10, pady=10, sticky='w')
         
     def _clicked_vid(self):
         self.vidName = filedialog.askopenfilenames(initialdir=VID_DEFAULT_FOLDER)
@@ -524,15 +527,20 @@ class InputSelectorGUI:
         for element in self.vidElement.values():
             element.destroy()
 
-        self.imgElement['lblFolder'] = ttk.Label(self.root, text="Image's folder: ",  width = 15)
-        self.imgElement['lblFolder'].grid(row=2, column=0, padx=10, pady=10)
+        self.imgElement['lblFolder'] = ttk.Label(self.root, text="Image's folder: ")
+        self.imgElement['lblFolder'].grid(row=2, column=1, padx=10, pady=10, sticky='w')
         self.imgElement['lblFolderName'] = ttk.Label(self.root, text="Waiting for the selection", wraplength=220)
-        self.imgElement['lblFolderName'].grid(row=2, column=1, columnspan=2, padx=10, pady=30)
+        self.imgElement['lblFolderName'].grid(row=2, column=2, padx=10, pady=30, sticky='w')
 
         self.imgElement['btnStart'] = ttk.Button(self.root, text="Select start frame", command=self._clicked_start_img)
-        self.imgElement['btnStart'].grid(row=3, column=1,  padx=10, pady=10)
+        self.imgElement['btnStart'].grid(row=3, column=1,  padx=10, pady=10, sticky='w')
+        self.imgElement['lblStartImg'] = ttk.Label(self.root, text=self.startImgName)
+        self.imgElement['lblStartImg'].grid(row=3, column=2, padx=10, pady=10, sticky='w')
+
         self.imgElement['btnEnd'] = ttk.Button(self.root, text="Select end frame", command=self._clicked_end_img)
-        self.imgElement['btnEnd'].grid(row=4, column=1,  padx=10, pady=10)
+        self.imgElement['btnEnd'].grid(row=4, column=1,  padx=10, pady=10, sticky='w')
+        self.imgElement['lblEndImg'] = ttk.Label(self.root, text=self.endImgName)
+        self.imgElement['lblEndImg'].grid(row=4, column=2, padx=10, pady=10, sticky='w')
         
     def _clicked_start_img(self):
         startImgFullPath = filedialog.askopenfilenames(
@@ -551,8 +559,7 @@ class InputSelectorGUI:
         self.imgSelectFolder = self.startFolder
         self.imgElement['lblFolderName'].config(text=self.imgSelectFolder)
 
-        self.imgElement['lblStartImg'] = ttk.Label(self.root, text=self.startImgName)
-        self.imgElement['lblStartImg'].grid(row=3, column=2, padx=10, pady=10)
+        self.imgElement['lblStartImg'].config(text=self.startImgName)
 
     def _clicked_end_img(self):
         endImgFullPath = filedialog.askopenfilenames(
@@ -573,16 +580,127 @@ class InputSelectorGUI:
         self.imgSelectFolder = self.endFolder
         self.imgElement['lblFolderName'].config(text=self.imgSelectFolder)
 
-        self.imgElement['lblEndImg'] = ttk.Label(self.root, text=self.endImgName)
-        self.imgElement['lblEndImg'].grid(row=4, column=2, padx=10, pady=10)
+        self.imgElement['lblEndImg'].config(text=self.endImgName)
+
+
+class PostProcessingSelectorGUI:
+    def __init__(self, root):
+        self.root = root
+        self.output_type = "dot"
+        self.show_threshold = 0.0
+        self.top_num = 1
+
+        self.outputTypeLabel = ttk.Label(self.root, text="Output Type:", width = 15)
+        self.outputTypeLabel.grid(row=5, column=0, padx=10, pady=10)
+
+        self.selectedOption = tk.IntVar(value=2)
+
+        self.dotLabel = ttk.Radiobutton(self.root, 
+                                        text='dot output', 
+                                        variable=self.selectedOption,
+                                        value=2, 
+                                        command=self.select_dot)
+        self.dotLabel.grid(row=5, column=1, padx=10, pady=10, sticky="w")
+        
+        self.bboxLabel = ttk.Radiobutton(self.root, 
+                                        text='bbox output', 
+                                        variable=self.selectedOption,
+                                        value=1, 
+                                        command=self.select_bbox)
+        self.bboxLabel.grid(row=5, column=2, padx=10, pady=10, sticky="w")
+
+        self.showThresholdLabel = ttk.Label(self.root, text="Threshold:", width=10)
+        self.showThresholdLabel.grid(row=6, column=1, padx=10, pady=10, sticky='w')
 
+        self.showThresholdVar = tk.StringVar(value="0")
+        self.showThresholdVar.trace_add('write', self.update_show_threshold)
+        self.showThresholdEntry = ttk.Entry(self.root, textvariable=self.showThresholdVar, width=5)
+        self.showThresholdEntry.grid(row=6, column=2, padx=10, pady=10, sticky='w')
 
-class ModelAndInputSelectorGUI:
+        self.getTopNumLabel = ttk.Label(self.root, text="Top Num:", width=10)
+        self.getTopNumLabel.grid(row=7, column=1, padx=10, pady=10, sticky='w')
+
+        self.getTopNumVar = tk.StringVar(value="1")
+        self.getTopNumVar.trace_add('write', self.update_top_num)
+        self.getTopNumEntry = ttk.Entry(self.root, textvariable=self.getTopNumVar, width=5)
+        self.getTopNumEntry.grid(row=7, column=2, padx=10, pady=10, sticky='w')
+        
+        self.select_dot()
+
+    def select_dot(self):
+        self.selectedOption.set(2)
+        self.output_type = "dot"
+
+    def select_bbox(self):
+        self.selectedOption.set(1)
+        self.output_type = "bbox"
+
+    def get_post_processing(self):
+        if self.output_type == "dot":
+            return PostProcessing(get_top_num = self.top_num)
+        elif self.output_type == "bbox":
+            return bbox_post_processing(self.top_num)
+        else:
+            raise ValueError(f"Unknown output type: {self.output_type}")
+
+    def update_show_threshold(self, *args):
+        value = float(self.showThresholdVar.get())
+        self.show_threshold = min(max(value, 0.0), 1.0)  # 确保在 [0.0, 1.0] 范围内
+        
+    def update_top_num(self, *args):
+        value = int(self.getTopNumVar.get())
+        self.top_num = max(value, 1)  # 确保 top_num 至少为 1
+
+
+
+
+class DeviceSelectorGUI:
     def __init__(self, root):
         self.root = root
+        self.device = "cpu"
 
-        windowHeight = 350
-        windowWidth = 400
+        self.deviceLabel = ttk.Label(self.root, text="Select Device:", width=15)
+        self.deviceLabel.grid(row=8, column=0, padx=10, pady=10)
+
+        self.selectedOption = tk.IntVar(value=1)
+
+        if torch.cuda.is_available():
+            self.selectedOption.set(2)
+            self.device = "cuda"
+
+        self.cpuLabel = ttk.Radiobutton(self.root, 
+                                        text='CPU', 
+                                        variable=self.selectedOption,
+                                        value=1, 
+                                        command=self.select_cpu)
+        self.cpuLabel.grid(row=8, column=1, padx=10, pady=10, sticky="w")
+        
+        self.gpuLabel = ttk.Radiobutton(self.root, 
+                                        text='GPU', 
+                                        variable=self.selectedOption,
+                                        value=2, 
+                                        command=self.select_gpu)
+        self.gpuLabel.grid(row=8, column=2, padx=10, pady=10, sticky="w")
+
+    def select_cpu(self):
+        self.selectedOption.set(1)
+        self.device = "cpu"
+
+    def select_gpu(self):
+        if torch.cuda.is_available():
+            self.selectedOption.set(2)
+            self.device = "cuda"
+        else:
+            messagebox.showinfo("Message title", "CUDA is not available. Please select CPU.")
+            self.select_cpu()
+
+
+class XTTMP_GUI:
+    def __init__(self, root):
+        self.root = root
+
+        windowHeight = 550
+        windowWidth = 510
         
         startHeight = (root.winfo_screenheight() - windowHeight) // 2
         startWidth = (root.winfo_screenwidth() - windowWidth) // 2
@@ -592,13 +710,16 @@ class ModelAndInputSelectorGUI:
         self._set_window_icon()
         
         self.objModelSelector = ModelSelectorGUI(root)
+        self.objPostProcessingSelector = PostProcessingSelectorGUI(root)
         self.objInputSelector = InputSelectorGUI(root)
+        self.objDeviceSelector = DeviceSelectorGUI(root)
+        
         
-        self.btnRun = ttk.Button(self.root, text="Run", command=self._run)
-        self.btnRun.place(x = 20, y=300)
         self.btnStepping = ttk.Button(self.root, text="Stepping", command=self._stepping)
-        self.btnStepping.place(x = 20, y=270)
         self.isStepping = False
+        self.btnStepping.grid(row=9, column=2, padx=10, pady=10, sticky='e')
+        self.btnRun = ttk.Button(self.root, text="Run", command=self._run)
+        self.btnRun.grid(row=10, column=2, padx=10, pady=10, sticky='e')
 
     def create_gui(self):
         self.objModelSelector.create_gui(ALL_MODEL)
@@ -607,9 +728,13 @@ class ModelAndInputSelectorGUI:
         self.root.mainloop()
 
         if self.objInputSelector.selectedOption.get() == 1:  
-            return self.modelName, self.vidName, None, self.isStepping
+            return (self.modelName, self.vidName, None, self.isStepping, 
+                    self.objDeviceSelector.device, self.objPostProcessingSelector.get_post_processing(),
+                    self.objPostProcessingSelector.show_threshold)
         elif self.objInputSelector.selectedOption.get() == 2:
-            return self.modelName, self.startImgName, self.endImgName, self.isStepping
+            return (self.modelName, self.startImgName, self.endImgName, self.isStepping,
+                     self.objDeviceSelector.device, self.objPostProcessingSelector.get_post_processing(),
+                    self.objPostProcessingSelector.show_threshold)
 
     def _run(self):
         self.modelName = self.objModelSelector.modelCombobox.get()

{xttmp-2.3.0.1 → xttmp-2.3.0.3/src/xttmp.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: xttmp
-Version: 2.3.0.1
+Version: 2.3.0.3
 Summary: eXtremely Tiny Target - Motion Perception
 Author-email: Shawn MX <mingshuoxu@hotmail.com>
 Project-URL: Homepage, https://github.com/MingshuoXu/Small-Target-Motion-Detectors
@@ -16,15 +16,16 @@ Classifier: Programming Language :: Python :: 3.9
 Classifier: Programming Language :: Python :: 3.10
 Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
-Classifier: Topic :: Scientific/Engineering :: Image Recognition
+Classifier: Topic :: Scientific/Engineering :: Artificial Intelligence
 Requires-Python: >=3.8
 Description-Content-Type: text/markdown
 License-File: LICENSE
 Requires-Dist: matplotlib
 Requires-Dist: opencv-python
 Requires-Dist: scipy
-Requires-Dist: torch>=2.5.0
-Requires-Dist: torchvision>=0.20.0
+Provides-Extra: torch
+Requires-Dist: torch>=2.5.0; extra == "torch"
+Requires-Dist: torchvision>=0.20.0; extra == "torch"
 Dynamic: license-file
 
 # Small Target Motion Detectors, Version 2.3 (XTT-MP: Extremely Tiny Target - Motion Perception)
@@ -66,8 +67,18 @@ Built with modularity and extensibility in mind, XTT-MP provides a robust suite
 - After `pip install xttmp`, use the installed code and bring your own input data, or run from a repository checkout to access the bundled examples.
 
 ### Via PyPI
+#### CPU
+```bash
+pip install xttmp[torch]
+```
+
+#### NVIDIA GPU (CUDA 12.6)
+```bash
+pip install torch torchvision --index-url https://download.pytorch.org/whl/cu126
+```
+
+### Running the GUI Demo
 ```bash
-pip install xttmp
 xttmp_gui
 ```
 

{xttmp-2.3.0.1 → xttmp-2.3.0.3}/src/xttmp.egg-info/requires.txt

@@ -1,5 +1,7 @@
 matplotlib
 opencv-python
 scipy
+
+[torch]
 torch>=2.5.0
 torchvision>=0.20.0