Tutorial 6: Find false negatives of object detection

In object detection applications, it can happen that the detector does not detect an object because it did not see any examples of this or similar objects yet. This is especially a problem in warehouse and retail applications, as new products get added to the shelves every day. These new products with new appearence often have a low so-called objectness score. Currently, finding such negative samples requires a lot of manual labour, especially when having 1000s of new images coming in every day. Lightly can make this work easier in a two-step approach:

  1. A human finds one false negative and adds it to the missing examples.

2. Lightly finds all similar images which are also false negatives. Thus it can propose to directly also add them as missing examples.

If there are e.g. 9 similar images for each missing example found by a human, Lightly can speed up the process by a factor of 10.

What you will learn

  • How to use an object detection model to crop objects out of a full image and save them as images.

  • How to save the object detection scores as metadata in a Lightly format.

  • How to upload the cropped images with their corresponding metadata to the webapp.

  • How to use the webapp to find false negatives and similar examples easily.


You can use your own dataset or the one we provide for this tutorial. The dataset we will use is the SKU110k dataset showing store shelves. The tutorial is computationally expensive if you run it on the full 110k images of products, thus we recommend running it on a subset of the dataset. E.g. copy 100 images from one folder to a new folder and use the path to the latter folder as input. Alternatively, run it on your own dataset.

For this tutorial the lightly pip package needs to be installed:

# Install lightly as a pip package
pip install lightly


The steps of this tutorial are quite straightforward:
  1. Define the dataset in form of a LightlyDataset of torch tensors. You need to provide the path to the dataset images.

  2. Define a pretrained object detection model. We use the retina net trained on COCO 2017. As it was not pretrained on a retail dataset, its performance is not state-of-the art. Nonetheless, it is sufficient for this tutorial and very easy to use.

  3. Generate predictions for each image in the dataset.

  4. Use the bounding boxes of the object predictions to crop the objects out of the full images and save them in the output directory.

  5. Extract the objectness scores and save them as custom metadata in a .json file.

  6. Use the lightly-magic command to upload the cropped images, their embeddings and the objectness scores.

  7. In the Lightly Webapp: Configure the objectness score as custom metadata.

  8. In the Lightly Webapp: Sort the images in the explore view by increasing objectness score. This allows to easily find missing examples / false positives and similar images to them.

For 100 input images with 150 predicted objects on each image, the tutorial runs in about 30 minutes on a Laptop CPU.

import torch
import torchvision
from torch.utils.data import DataLoader
from tqdm import tqdm

from lightly.active_learning.utils import BoundingBox
from lightly.data import LightlyDataset
from lightly.utils import save_custom_metadata
from lightly.utils.cropping.crop_image_by_bounding_boxes import crop_dataset_by_bounding_boxes_and_save

BASE_PATH = "path/to/dataset/"
DATASET_PATH = os.path.join(BASE_PATH, "images")  # the path were the full images are found
OUTPUT_DIR = os.path.join(BASE_PATH, "cropped_images") # the path where the cropped images will be saved
# the file where the objectness scores will be saved
METADATA_OUTPUT_FILE = os.path.join(BASE_PATH, "cropped_images_objectness_scores.json")

''' 1. Define the dataset'''
x_size = 2048
y_size = 2048
transform = torchvision.transforms.Compose([
    torchvision.transforms.Resize((x_size, y_size)),
            mean=[0.485, 0.456, 0.406],
            std=[0.229, 0.224, 0.225])
dataset = LightlyDataset(DATASET_PATH, transform=transform)

''' 2. Define the pretrained object detection model'''
model = torchvision.models.detection.retinanet_resnet50_fpn(pretrained=True)

''' 3. Predict with the model on the dataset '''
dataloader = DataLoader(dataset, batch_size=2)
predictions = []
with torch.no_grad():
    for x, _, _ in tqdm(dataloader):
        pred = model(x)

predictions = [i for sublist in predictions for i in sublist]

''' 4. Save the cropped objects '''
class_indices_list_list = [list(prediction["labels"]) for prediction in predictions]
bounding_boxes_list_list = []
for prediction in predictions:
    bounding_boxes_list = []
    for box in prediction["boxes"]:
        x0 = box[0] / x_size
        y0 = box[1] / y_size
        x1 = box[2] / x_size
        y1 = box[3] / y_size
        bounding_boxes_list.append(BoundingBox(x0, y0, x1, y1))

cropped_images_list_list = crop_dataset_by_bounding_boxes_and_save(dataset, OUTPUT_DIR,
    bounding_boxes_list_list, class_indices_list_list)

'''  5. Save the objectness scores as metadata '''
objectness_scores_list_list = [list(prediction["scores"]) for prediction in predictions]
metadata_list = []
for cropped_images_list, objectness_scores_list in zip(cropped_images_list_list, objectness_scores_list_list):
    for cropped_images_filename, objectness_score in zip(cropped_images_list, objectness_scores_list):
        metadata = {"objectness_score": float(objectness_score)}
        metadata_list.append((cropped_images_filename, metadata))
save_custom_metadata(METADATA_OUTPUT_FILE, metadata_list)

''' 6. Tell the lightly CLI command '''
cli_command = f"lightly-magic input_dir={OUTPUT_DIR} new_dataset_name=SKU_110k_val_cropped trainer.max_epochs=0 "
cli_command += f"custom_metadata={METADATA_OUTPUT_FILE} token=MY_TOKEN"
print(f"Upload the images and custom metadata with the following CLI command:")

6. Adapt this command to include your Lightly webapp token and run it in a terminal. It will embed the images with a pretrained model, create a new dataset in the Lightly webapp and upload the images, embeddings and metadata to it. You can also change some arguments, e.g. to train a better embedding model instead of relying on a pretrained one. For more information, head to Command-line tool.

Terminal output of lightly-magic command.

7. Once the cropped images, embeddings and metadata are uploaded, you can use the Lightly Webapp to configure the Objectness Score as metadata. This is done in the Configurator view.

Configuration of Objectness Score as metadata

8. Now you can switch to the explore view and select to sort by the Objectness Score in ascending order. You see that many images show objects despite having a low objectness score, thus they are false negatives / missing examples.

Explore view of images sorted by ascending objectness score

When clicking on one them, you see that it has a low objectness score of only 0.05, despite showing an object, thus it is false negative. Similar images which are also false negatives are shown as well. Thus all of them can be added directly to the list of missing examples, instead of finding and adding all of them by hand.

Detail view of a missing example together with similar samples

You can select/unselect an image by clicking on the checkbox in its top-right corner. Multiple images in a row can be selected with shift-click, all images with CTRL-A. You can view all currently selected images by toggling “Only show selected” in the top.

View of all selected images

If you are satisfied with your current selection, click on the green double arrow buttom in the bottom right to create a new Tag from the current selection. You can call this tag e.g. “missing_examples”

To download these images or their filenames head to the Download View. You can download from there directly or using the CLI command.

Total running time of the script: ( 0 minutes 0.000 seconds)

Gallery generated by Sphinx-Gallery