simaiden / clothing-detection

Clothing detection using YOLOv3, RetinaNet, Faster RCNN in ModaNet and DeepFashion2 datasets.

Datasets

• DeepFashion2 dataset: https://github.com/switchablenorms/DeepFashion2

• ModaNet dataset: https://github.com/eBay/modanet

Models

• Faster RCNN, RetinaNet and Mask RCNN (only detection) trained with maskrcnn-benchmark https://github.com/facebookresearch/maskrcnn-benchmark/. To use this models please follow INSTALL instruccions in that repo and do the setup in the root folder of this repo. Not neccessary to use pytorch-nightly, you can use pytorch 1.2 instead.

• YOLOv3 trained with Darknet framework: https://github.com/AlexeyAB/darknet

• TridenNet trained with simpledet framework https://github.com/TuSimple/simpledet

• To do inference use a pytorch implementation of YOLOv3: https://github.com/eriklindernoren/PyTorch-YOLOv3.

• All the models trained with Resnet50 backbone, except YOLOv3 with Darknet53

Weights

All weights and config files are in https://drive.google.com/drive/folders/1jXZZc5pp2OJCtmQYelzDgPzyuraAdxXP?usp=sharing

Using

• Use new_image_demo.py , and choose dataset, and model.
• Use YOLOv3Predictor class for YOLOv3 and Predictor class for Faster and RetinaNet and Mask.

Coming soon

• Update use of retrieval.

clothing-detection's Issues

AUC metrics on test set

What AUC scores did you get on your test set for YOLOV3 MODANET, and what test set did you use? I'm trying to replicate your training and testing.

Annotation of the dataset

Hi, just wondering how did you train your yolov3 model to for the classification? did you convert the json files to txt files for training?

Cannot find yolov3-modanet_last.weights

Getting the error

Traceback (most recent call last):
  File "new_image_demo.py", line 62, in <module>
    detectron = YOLOv3Predictor(params=yolo_params)
  File "/home/ubuntu/server/Clothing-Detection/predictors/YOLOv3.py", line 22, in __init__
    self.model = self.load_model()
  File "/home/ubuntu/server/Clothing-Detection/predictors/YOLOv3.py", line 59, in load_model
    model.load_darknet_weights(self.params['weights_path'])
  File "/home/ubuntu/server/Clothing-Detection/yolo/utils/models.py", line 312, in load_darknet_weights
    with open(weights_path, "rb") as f:
FileNotFoundError: [Errno 2] No such file or directory: 'yolo/weights/yolov3-modanet_last.weights'

it seems like download_weights.sh doesn't include the yolov3-modanet_last.weights file any ideas ?

Is there any way to use segmentation?

Hi,

Is there any segmentation model in this repo? I couldn't find out it.
Or, is it possible to adapt this application to the segmentation problem?

Thank you!

I want to know the class information of maskrcnn_banchmark

Thank you for your project.
Actually I am using your maskrcnn_benchmark to detect clothes.
Detection works fine.
However, class matching is not good. Can you tell me the class information of maskrcnn_benchmark?

DeepFashion2 YOLOv3 detect not very well?

I have test DeppFashion2 with yolov3 it seams not very well ? Do you have a new yolov3 weights for DeepFashion2? thanks for your work

Missing yolov3-df2_15000.weights ?

hi,

can anyone provide link to this file yolov3-df2_15000.weights

Thanks

Confidence scores are too low despite while using deepfashion2 validation set

I've used the YOLOv3 model for Deepfashion2 with the validation set given by deepfashion2 and my scores are so low despite using 5,000 and 10,000 images. The mean average precision is low as well. Can I ask how did you come up with your own result for coco evaluation?

No weights in gdrive link

Please upload it asap , my final year project is highly dependent on this

can not load model

size mismatch for roi_heads.box.predictor.cls_score.weight: copying a param with shape torch.Size([14, 1024]) from checkpoint, the shape in current model is torch.Size([81, 1024]).
size mismatch for roi_heads.box.predictor.cls_score.bias: copying a param with shape torch.Size([14]) from checkpoint, the shape in current model is torch.Size([81]).
size mismatch for roi_heads.box.predictor.bbox_pred.weight: copying a param with shape torch.Size([56, 1024]) from checkpoint, the shape in current model is torch.Size([324, 1024]).
size mismatch for roi_heads.box.predictor.bbox_pred.bias: copying a param with shape torch.Size([56]) from checkpoint, the shape in current model is torch.Size([324]).

Cannot detect anything with the provided weights and model

Hey there, I want to say sorry beforehand that I ask this (possibly) stupid question, but I am new to object detection and all it's nooks and crevices. I wanted to ask you if you could help me with this problem? I have modified the webcam_demo.py file so that it loads the yolov3-modanet.cfg model, the modanet.names classes, and yolov3.weights weights. I have followed the tutorial in this video: https://www.youtube.com/watch?v=yWwzFnAnrLM&ab_channel=AlSangam but to no avail, perhaps the weight file is different?

Here's my modified code inside the webcam_demo.py file.

from __future__ import division

from yolo.utils.models import *
from yolo.utils.utils import *
from yolo.utils.datasets import *
from predictors.YOLOv3 import YOLOv3Predictor

import torch
from torch.utils.data import DataLoader
from torchvision import datasets
from torch.autograd import Variable

import matplotlib.pyplot as plt
import cv2

from yolo.utils.utils2 import load_model, cv_img_to_tensor

device = torch.device("cuda" if torch.cuda.is_available() else "cpu")

# params = \
# {
#   "model_def" : "yolo/df2cfg/yolov3-df2.cfg",
#   "weights_path" : "yolo/weights/yolov3.weights",
#   "class_path":"yolo/df2cfg/df2.names",
#   "conf_thres" : 0.25,
#   "nms_thres" :0.4,
#   "img_size" : 416,
#   "device" : device
# }

params = \
{
  "model_def" : "yolo/modanetcfg/yolov3-modanet.cfg",
  "weights_path" : "yolo/weights/yolov3.weights",
  "class_path":"yolo/modanetcfg/modanet.names",
  "conf_thres" : 0.25,
  "nms_thres" :0.4,
  "img_size" : 416,
  "device" : device
}


classes = load_classes(params['class_path'])
Tensor = torch.cuda.FloatTensor if torch.cuda.is_available() else torch.FloatTensor
cmap = plt.get_cmap("tab20b")
colors = np.array([cmap(i) for i in np.linspace(0, 1, 20)])
np.random.shuffle(colors)

yolov3 = YOLOv3Predictor(params=params)
model = yolov3.load_model()

cap = cv2.VideoCapture(0)

while(True):
  # img = cv2.imread('tests/amazon.jpg')
  _, frame = cap.read()
  img = frame.copy()
  x , _,_ = cv_img_to_tensor(img)
  x.to(device)   
  
  # Get detections
  with torch.no_grad():
    input_img = Variable(x.type(Tensor))  
    detections = model(input_img)
    detections = non_max_suppression(detections, params['conf_thres'], params['nms_thres'])
  
  if detections[0] is not None:
    # Rescale boxes to original image
    detections = rescale_boxes(detections[0], params['img_size'], img.shape[:2])
    unique_labels = detections[:, -1].cpu().unique()
    n_cls_preds = len(unique_labels)
    # bbox_colors = random.sample(colors, n_cls_preds , seed)
    bbox_colors = colors[:n_cls_preds]

    for x1, y1, x2, y2, conf, cls_conf, cls_pred in detections:
      print("\t+ Label: %s, Conf: %.5f" % (classes[int(cls_pred)], cls_conf.item()))
      x1, y1, x2, y2 = int(x1), int(y1), int(x2), int(y2)

      color = bbox_colors[int(np.where(unique_labels == int(cls_pred))[0])]
      color = tuple(c*255 for c in color)
      color = (color[2],color[1],color[0])
      cv2.rectangle(frame, (x1,y1), (x2,y2), color=color, thickness=3)
      # cv2.rectangle(img, (x1,y1), (x2,y2), color=color, thickness=3)
      # print(int(cls_pred))
      font = cv2.FONT_HERSHEY_SIMPLEX
      text =  "%s conf: %.3f" % (classes[int(cls_pred)] ,cls_conf.item())
      cv2.rectangle(frame,(x1-2,y1-25), (int(x1 + 8.5*len(text)),y1) , color=color, thickness=-1)
      cv2.putText(frame,text,(x1,y1-5), font, 0.5, color=(255,255,255), thickness=1, lineType=cv2.LINE_AA)
      # cv2.rectangle(img,(x1-2,y1-25), (int(x1 + 8.5*len(text)),y1) , color=color, thickness=-1)
      # cv2.putText(img,text,(x1,y1-5), font, 0.5, color=(255,255,255), thickness=1, lineType=cv2.LINE_AA)
            
  cv2.imshow('Detections',frame)
  # cv2.imshow('Detections',img)
  if cv2.waitKey(30) & 0xFF == ord('q'):
      break

I have tested the model with my webcam and the test images but to no avail, no bounding boxes are created and the model does not seem to detect anything at all! If I change the model to darknet53.conv.74 weights file, there are bounding boxes everywhere! Like so:

Can you give me a hint or a solution to this problem, or the thing that might went wrong? Thank you so much! I need this for my personal research :( A timely response will be appreciated. This GitHub Issue is my last hope in finishing my research on time...

Thank you so much beforehand, for your possible help and this very helpful repository! 🚀

cannot import name '_C' from 'maskrcnn_benchmark'

Hello,
I am new to this ML thing
i have done setup for maskrcnn_benchmark i dont have GPU.
all the time i am getting "ImportError: cannot import name '_C' from 'maskrcnn_benchmark'" error
Please Help

error loading state_dict

the model architecture in config file and the weight file seems to have different shape.
I am getting this error:
size mismatch for roi_heads.box.predictor.cls_score.weight: copying a param with shape torch.Size([14, 1024]) from checkpoint, the shape in current model is torch.Size([81, 1024]).
size mismatch for roi_heads.box.predictor.cls_score.bias: copying a param with shape torch.Size([14]) from checkpoint, the shape in current model is torch.Size([81]).
size mismatch for roi_heads.box.predictor.bbox_pred.weight: copying a param with shape torch.Size([56, 1024]) from checkpoint, the shape in current model is torch.Size([324, 1024]).
size mismatch for roi_heads.box.predictor.bbox_pred.bias: copying a param with shape torch.Size([56]) from checkpoint, the shape in current model is torch.Size([324]).

Dataset link

Can you please share the link for ModaNet dataset download

does not detect well

Hello,

I downloaded your code and tried to run it on an image from the web but it does not work. It gives me thousands of bounding boxes. Are you sure from your training on DeepFashion2? and are you sure of your non maximum suppression function because i found others on the web but still no results.

Thanks in advance