import torch import os import streamlit as st import matplotlib.pyplot as plt import pandas as pd import numpy as np import plotly.express as px import pickle import random from utils import check_password from PIL import Image from transformers import YolosFeatureExtractor, YolosForObjectDetection from torchvision.transforms import ToTensor, ToPILImage from annotated_text import annotated_text from st_pages import add_indentation #add_indentation() st.set_page_config(layout="wide") @st.cache_data(ttl=3600, show_spinner=False) def load_model(feature_extractor_url, model_url): feature_extractor_ = YolosFeatureExtractor.from_pretrained(feature_extractor_url) model_ = YolosForObjectDetection.from_pretrained(model_url) return feature_extractor_, model_ def rgb_to_hex(rgb): """Converts an RGB tuple to an HTML-style Hex string.""" hex_color = "#{:02x}{:02x}{:02x}".format(int(rgb[0] * 255), int(rgb[1] * 255), int(rgb[2] * 255)) return hex_color ## CODE TO CLEAN IMAGES def fix_channels(t): if len(t.shape) == 2: return ToPILImage()(torch.stack([t for i in (0, 0, 0)])) if t.shape[0] == 4: return ToPILImage()(t[:3]) if t.shape[0] == 1: return ToPILImage()(torch.stack([t[0] for i in (0, 0, 0)])) return ToPILImage()(t) ## CODE FOR PLOTS WITH BOUNDING BOXES COLORS = [[0.000, 0.447, 0.741], [0.850, 0.325, 0.098], [0.929, 0.694, 0.125], [0.494, 0.184, 0.556], [0.466, 0.674, 0.188], [0.301, 0.745, 0.933]] def idx_to_text(i): if i in list(dict_cats_final.keys()): return dict_cats_final[i.item()] else: return False # for output bounding box post-processing def box_cxcywh_to_xyxy(x): x_c, y_c, w, h = x.unbind(1) b = [(x_c - 0.5 * w), (y_c - 0.5 * h), (x_c + 0.5 * w), (y_c + 0.5 * h)] return torch.stack(b, dim=1) def rescale_bboxes(out_bbox, size): img_w, img_h = size b = box_cxcywh_to_xyxy(out_bbox) b = b * torch.tensor([img_w, img_h, img_w, img_h], dtype=torch.float32) return b def plot_results(pil_img, prob, boxes): fig = plt.figure(figsize=(16,10)) plt.imshow(pil_img) ax = plt.gca() colors = COLORS * 100 colors_used = [] for p, (xmin, ymin, xmax, ymax), c in zip(prob, boxes.tolist(), colors): cl = p.argmax() p_max = p.max().detach().numpy() if idx_to_text(cl) is False: pass else: colors_used.append(rgb_to_hex(c)) ax.add_patch(plt.Rectangle((xmin, ymin), xmax - xmin, ymax - ymin, fill=False, color=c, linewidth=3)) ax.text(xmin, ymin, f"{idx_to_text(cl)}", fontsize=10, bbox=dict(facecolor=c, alpha=0.8)) plt.axis('off') plt.savefig("results_od.png", bbox_inches ="tight") plt.show() st.image("results_od.png") return colors_used def return_probas(outputs, threshold): probas = outputs.logits.softmax(-1)[0, :, :-1] probas = probas[:][:,list(dict_cats_final.keys())] keep = probas.max(-1).values > threshold return probas, keep def visualize_probas(probas, threshold, colors): label_df = pd.DataFrame({"label":probas.max(-1).indices.detach().numpy(), "proba":probas.max(-1).values.detach().numpy()}) cats_dict = dict(zip(np.arange(0,len(cats)),cats)) label_df["label"] = label_df["label"].map(cats_dict) top_label_df = label_df.loc[label_df["proba"]>threshold].round(2) top_label_df["colors"] = colors top_label_df.sort_values(by=["proba"], ascending=False, inplace=True) #st.dataframe(top_label_df.drop(columns=["colors"])) mode_func = lambda x: x.mode().iloc[0] top_label_df_agg = top_label_df.groupby("label").agg({"proba":"mean", "colors":mode_func}) top_label_df_agg = top_label_df_agg.reset_index().sort_values(by=["proba"], ascending=False) top_label_df_agg.columns = ["Item","Score","Colors"] color_map = dict(zip(top_label_df_agg["Item"].to_list(), top_label_df_agg["Colors"].to_list())) fig = px.bar(top_label_df_agg, y='Item', x='Score', color="Item", title="Probability scores") st.plotly_chart(fig, use_container_width=True) cats = ['shirt, blouse', 'top, t-shirt, sweatshirt', 'sweater', 'cardigan', 'jacket', 'vest', 'pants', 'shorts', 'skirt', 'coat', 'dress', 'jumpsuit', 'cape', 'glasses', 'hat', 'headband, head covering, hair accessory', 'tie', 'glove', 'watch', 'belt', 'leg warmer', 'tights, stockings', 'sock', 'shoe', 'bag, wallet', 'scarf', 'umbrella', 'hood', 'collar', 'lapel', 'epaulette', 'sleeve', 'pocket', 'neckline', 'buckle', 'zipper', 'applique', 'bead', 'bow', 'flower', 'fringe', 'ribbon', 'rivet', 'ruffle', 'sequin', 'tassel'] ###################################################################################################################################### if check_password(): st.markdown("# Object Detection 📹") st.markdown("### What is Object Detection ?") #st.markdown("""Object detection involves **identifying** and **locating objects** within an image or video frame through bounding boxes. """) st.info("""Object Detection is a computer vision task in which the goal is to **detect** and **locate objects** of interest in an image or video. The task involves identifying the position and boundaries of objects (or **bounding boxes**) in an image, and classifying the objects into different categories.""") st.markdown("Here is an example of Object Detection for Traffic Analysis.") #image_od = Image.open('images/od_2.png') #st.image(image_od, width=600) st.video(data='https://www.youtube.com/watch?v=PVCGDoTZHaI') st.markdown(" ") st.markdown("""Common applications of Object Detection include: - **Autonomous Vehicles** :car: : Object detection is crucial for self-driving cars to track pedestrians, cyclists, other vehicles, and obstacles on the road. - **Retail** 🏬 : Implementing smart shelves and checkout systems that use object detection to track inventory and monitor stock levels. - **Healthcare** 👨‍⚕️: Detecting and tracking anomalies in medical images, such as tumors or abnormalities, for diagnostic purposes or prevention. - **Manufacturing** 🏭: Quality control on production lines by detecting defects or irregularities in manufactured products. Ensuring workplace safety by monitoring the movement of workers and equipment. """) ############################# USE CASE ############################# st.markdown(" ") st.divider() st.markdown("# Fashion Object Detection 👗") # st.info("""This use case showcases the application of **Object detection** to detect clothing items/features on images.
# The images used were gathered from Dior's""") st.info("""**Object detection models** can very valuable for fashion retailers wishing to improve customer experience. They can provide, for example, **product recognition**, **visual search** and even **virtual try-ons**.""") st.markdown("In this use case, we are going to show an object detection model that as able to identify and locate different articles of clothings on fashion show images.") st.markdown(" ") st.markdown(" ") # images_dior = [os.path.join("data/dior_show/images",url) for url in os.listdir("data/dior_show/images") if url != "results"] # columns_img = st.columns(4) # for img, col in zip(images_dior,columns_img): # with col: # st.image(img) _, col, _ = st.columns([0.1,0.8,0.1]) with col: st.image("images/fashion_od2.png") st.markdown(" ") st.markdown(" ") st.markdown("### About the model 📚") st.markdown("""The object detection model was trained to **detect specific clothing items** on images.
Below is a list of the 46 different types of clothing items the model can identify and locate.""", unsafe_allow_html=True) colors = ["#8ef", "#faa", "#afa", "#fea", "#8ef","#afa"]*7 + ["#8ef", "#faa", "#afa", "#fea"] cats_annotated = [(g,"","#afa") for g in cats] annotated_text([cats_annotated]) # st.markdown("""**Here are the 'objects' the model is able to detect**:
# 'shirt, blouse', 'top, t-shirt, sweatshirt', 'sweater', 'cardigan', 'jacket', 'vest', 'pants', 'shorts', 'skirt', # 'coat', 'dress', 'jumpsuit', 'cape', 'glasses', 'hat', 'headband, head covering, hair accessory', 'tie', 'glove', 'watch', # 'belt', 'leg warmer', 'tights, stockings', 'sock', 'shoe', 'bag, wallet', 'scarf', 'umbrella', 'hood', 'collar', 'lapel', # 'epaulette', 'sleeve', 'pocket', 'neckline', 'buckle', 'zipper', 'applique', 'bead', 'bow', 'flower', 'fringe', 'ribbon', 'rivet', # 'ruffle', 'sequin', 'tassel'""", unsafe_allow_html=True) st.markdown("Credits for the model: https://huggingface.co/valentinafeve/yolos-fashionpedia") st.markdown("") st.markdown("") ############## SELECT AN IMAGE ############### st.markdown("### Select an image 🖼️") st.markdown("""The images provided were taken from **Dior's 2020 Fall Women Fashion Show**""") image_ = None fashion_images_path = r"data/dior_show/images" list_images = os.listdir(fashion_images_path) image_name = st.selectbox("Choose an image", list_images) image_ = os.path.join(fashion_images_path, image_name) st.image(image_, width=300) # image_ = None # select_image_box = st.radio( # "**Select the image you wish to run the model on**", # ["Choose an existing image", "Load your own image"], # index=None,)# #label_visibility="collapsed") # if select_image_box == "Choose an existing image": # fashion_images_path = r"data/dior_show/images" # list_images = os.listdir(fashion_images_path) # image_ = st.selectbox("", list_images, label_visibility="collapsed") # if image_ is not None: # image_ = os.path.join(fashion_images_path,image_) # st.markdown("You've selected the following image:") # st.image(image_, width=300) # elif select_image_box == "Load your own image": # image_ = st.file_uploader("Load an image here", # key="OD_dior", type=['jpg','jpeg','png'], label_visibility="collapsed") # st.warning("""**Note**: The model tends to perform better with images of people/clothing items facing forward. # Choose this type of image if you want optimal results.""") # st.warning("""**Note:** The model was trained to detect clothing items on a single person. # If your image contains more than one person, the model won't detect the items of the other persons.""") # if image_ is not None: # st.image(Image.open(image_), width=300) st.markdown(" ") st.markdown(" ") ########## SELECT AN ELEMENT TO DETECT ################## dict_cats = dict(zip(np.arange(len(cats)), cats)) # st.markdown("#### Choose the elements you want to detect 👉") # # Select one or more elements to detect # container = st.container() # selected_options = None # all = st.checkbox("Select all") # if all: # selected_options = container.multiselect("**Select one or more items**", cats, cats) # else: # selected_options = container.multiselect("**Select one or more items**", cats) #cats = selected_options selected_options = cats dict_cats_final = {key:value for (key,value) in dict_cats.items() if value in selected_options} # st.markdown(" ") # st.markdown(" ") ############## SELECT A THRESHOLD ############### st.markdown("### Define a threshold for predictions 🔎") st.markdown("""In this section, you can select a threshold for the model's final predictions.
Objects that are given a lower score than the chosen threshold will be ignored in the final results""", unsafe_allow_html=True) st.info("""**Note**: Object detection models detect objects using bounding boxes as well as assign objects to specific classes. Each object is given a class based on a probability score computed by the model. A high probability signals that the model is confident in its prediction. On the contrary, a lower probability score signals a level of uncertainty.""") st.markdown(" ") #st.markdown("The images below are examples of probability scores given by object detection models for each element detected.") _, col, _ = st.columns([0.2,0.6,0.2]) with col: st.image("images/probability_od.png", caption="Examples of object detection with bounding boses and probability scores") st.markdown(" ") st.markdown("**Select a threshold** ") # st.warning("""**Note**: The threshold helps you decide how confident you want your model to be with its predictions. # Elements that are identified with a lower probability than the given threshold will be ignored in the final results.""") threshold = st.slider('**Select a threshold**', min_value=0.5, step=0.05, max_value=1.0, value=0.75, label_visibility="collapsed") # if threshold < 0.6: # st.error("""**Warning**: Selecting a low threshold (below 0.6) could lead the model to make errors and detect too many objects.""") st.write("You've selected a threshold at", threshold) st.markdown(" ") pickle_file_path = r"data/dior_show/results" ############# RUN MODEL ################ run_model = st.button("**Run the model**", type="primary") if run_model: if image_ != None and selected_options != None and threshold!= None: with st.spinner('Wait for it...'): ## SELECT IMAGE #st.write(image_) image = Image.open(image_) image = fix_channels(ToTensor()(image)) ## LOAD OBJECT DETECTION MODEL FEATURE_EXTRACTOR_PATH = "hustvl/yolos-small" MODEL_PATH = "valentinafeve/yolos-fashionpedia" # feature_extractor, model = load_model(FEATURE_EXTRACTOR_PATH, MODEL_PATH) # # RUN MODEL ON IMAGE # inputs = feature_extractor(images=image, return_tensors="pt") # outputs = model(**inputs) # Save results # pickle_file_path = r"data/dior_show/results" # image_name = image_.split('\\')[1][:5] # with open(os.path.join(pickle_file_path, f"{image_name}_results.pkl"), 'wb') as file: # pickle.dump(outputs, file) image_name = image_name[:5] path_load_pickle = os.path.join(pickle_file_path, f"{image_name}_results.pkl") with open(path_load_pickle, 'rb') as pickle_file: outputs = pickle.load(pickle_file) probas, keep = return_probas(outputs, threshold) st.markdown("#### See the results ☑️") # PLOT BOUNDING BOX AND BARS/PROBA col1, col2 = st.columns(2) with col1: st.markdown(" ") st.markdown("##### 1. Bounding box results") bboxes_scaled = rescale_bboxes(outputs.pred_boxes[0, keep].cpu(), image.size) colors_used = plot_results(image, probas[keep], bboxes_scaled) with col2: #st.markdown("**Probability scores**") if not any(keep.tolist()): st.error("""No objects were detected on the image. Decrease your threshold or choose differents items to detect.""") else: st.markdown(" ") st.markdown("##### 2. Probability score of each object") st.info("""**Note**: Some items might have been detected more than once on the image. For these items, we've computed the average probability score across all detections.""") visualize_probas(probas, threshold, colors_used) else: st.error("You must select an **image**, **elements to detect** and a **threshold** to run the model !")