Removing CutOff (WildChlamydia#15)

* Fixing readme and removing CutOff

---------

Co-authored-by: mvkuprashevich <[email protected]>
Co-authored-by: sweetdream779 <[email protected]>

.gitignore

@@ -80,3 +80,5 @@ Testing notebook.ipynb
 
 output/
 input/
+
+run.sh

README.md

@@ -263,12 +263,24 @@ python3 eval_pretrained.py \
 
 Supported dataset names: "utk", "imdb", "lagenda", "fairface", "adience".
 
+
+## Changelog
+
+15.08.20223 - 0.4.1dev
+
+### Added
+- Support for video streams, including YouTube URLs
+- Instructions and explanations for various export types.
+
+### Changed
+- Removed CutOff operation. It has been proven to be ineffective for inference time and quite costly at the same time. Now it is only used during training.
+
 ## ONNX and TensorRT export
 
-As of now (11.08.2023), while ONNX export is technically feasible, it is not advisable due to the poor performance of the resulting model with batch processing. 
-TensorRT export is impossible due to its lack of support for col2im.
+As of now (11.08.2023), while ONNX export is technically feasible, it is not advisable due to the poor performance of the resulting model with batch processing.
+**TensorRT** and **OpenVINO** export is impossible due to its lack of support for col2im.
 
-If you remain absolutely committed to utilizing ONNX export, you can refer to [these instructions](https://github.com/WildChlamydia/MiVOLO/issues/14#issuecomment-1675245889). 
+If you remain absolutely committed to utilizing ONNX export, you can refer to [these instructions](https://github.com/WildChlamydia/MiVOLO/issues/14#issuecomment-1675245889).
 
 The most highly recommended export method at present **is using TorchScript**. You can achieve this with a single line of code:
 ```python

mivolo/data/dataset/reader_age_gender.py

@@ -7,7 +7,7 @@
 import cv2
 import numpy as np
 from mivolo.data.data_reader import AnnotType, PictureInfo, get_all_files, read_csv_annotation_file
-from mivolo.data.misc import IOU, class_letterbox, cropout_black_parts
+from mivolo.data.misc import IOU, class_letterbox
 from timm.data.readers.reader import Reader
 from tqdm import tqdm
 
@@ -461,10 +461,9 @@ def _cropout_asced_objs(
 
         crop[aobj_ymin:aobj_ymax, aobj_xmin:aobj_xmax] = crop_out_color
 
-    crop, cropped_ratio = cropout_black_parts(crop, crop_round_tol)
-    if (
-        crop.shape[0] < min_person_size or crop.shape[1] < min_person_size
-    ) or cropped_ratio < min_person_aftercut_ratio:
+    # calc useful non-black area
+    remain_ratio = np.count_nonzero(crop) / (crop.shape[0] * crop.shape[1] * crop.shape[2])
+    if (crop.shape[0] < min_person_size or crop.shape[1] < min_person_size) or remain_ratio < min_person_aftercut_ratio:
         crop = None
         empty = True
 

mivolo/data/misc.py

@@ -22,27 +22,6 @@ def aggregate_votes_winsorized(ages, max_age_dist=6):
     return np.mean(ages)
 
 
-def cropout_black_parts(img, tol=0.3):
-    # Create a binary mask of zero pixels
-    zero_pixels_mask = np.all(img == 0, axis=2)
-    # Calculate the threshold for zero pixels in rows and columns
-    threshold = img.shape[0] - img.shape[0] * tol
-    # Calculate row sums and column sums of zero pixels mask
-    row_sums = np.sum(zero_pixels_mask, axis=1)
-    col_sums = np.sum(zero_pixels_mask, axis=0)
-    # Find the first and last rows with zero pixel sums above the threshold
-    start_row = np.argmin(row_sums > threshold)
-    end_row = img.shape[0] - np.argmin(row_sums[::-1] > threshold)
-    # Find the first and last columns with zero pixel sums above the threshold
-    start_col = np.argmin(col_sums > threshold)
-    end_col = img.shape[1] - np.argmin(col_sums[::-1] > threshold)
-    # Crop the image
-    cropped_img = img[start_row:end_row, start_col:end_col, :]
-    area = cropped_img.shape[0] * cropped_img.shape[1]
-    area_orig = img.shape[0] * img.shape[1]
-    return cropped_img, area / area_orig
-
-
 def natural_key(string_):
     """See http://www.codinghorror.com/blog/archives/001018.html"""
     return [int(s) if s.isdigit() else s for s in re.split(r"(\d+)", string_.lower())]

mivolo/predictor.py

@@ -56,7 +56,7 @@ def recognize_video(self, source: str) -> Generator:
             # add tr_persons and tr_faces to history
             for guid, data in cur_persons.items():
                 # not useful for tracking :)
-                if None not in data: 
+                if None not in data:
                     detected_objects_history[guid].append(data)
             for guid, data in cur_faces.items():
                 if None not in data:

mivolo/structures.py

@@ -6,7 +6,7 @@
 import cv2
 import numpy as np
 import torch
-from mivolo.data.misc import aggregate_votes_winsorized, assign_faces, box_iou, cropout_black_parts
+from mivolo.data.misc import aggregate_votes_winsorized, assign_faces, box_iou
 from ultralytics.yolo.engine.results import Results
 from ultralytics.yolo.utils.plotting import Annotator, colors
 
@@ -317,7 +317,7 @@ def _gather_tracking_result(
         person_ages = [r[0] for r in tracked_objects[pguid] if r[0] is not None] if pguid in tracked_objects else []
         person_genders = [r[1] for r in tracked_objects[pguid] if r[1] is not None] if pguid in tracked_objects else []
 
-        if not face_ages and not person_ages: # both empty
+        if not face_ages and not person_ages:  # both empty
             return None, None
 
         # You can play here with different aggregation strategies
@@ -434,7 +434,7 @@ def crop_object(
 
             obj_image[o_y1:o_y2, o_x1:o_x2] = 0
 
-        obj_image, remain_ratio = cropout_black_parts(obj_image, CROP_ROUND_RATE)
+        remain_ratio = np.count_nonzero(obj_image) / (obj_image.shape[0] * obj_image.shape[1] * obj_image.shape[2])
         if remain_ratio < MIN_PERSON_CROP_AFTERCUT_RATIO:
             return None
 

mivolo/version.py

@@ -1 +1 @@
-__version__ = "0.3.0dev"
+__version__ = "0.4.1dev"