Skip to main content

You should try to sign up to be a Mechanical Turk Requester ASAP. If you are unable to get an account, you can partner with a student who got one.

This assignment is before 11:59PM due on Tuesday, February 26, 2019.

Become a Requester : Assignment 5

In this homework assignment, you’re going to learn how to be a requester on Amazon Mechanical Turk. You should try to sign-up as a requester as soon as possible, because it is a multi-step process and each step can take time. If you’re unable to sign up as a requester yourself, we recommend partnering with another student who has successfully signed up. You will work in pairs.

Once you are a requester, you will be able to post work on MTurk and to pay workers to complete your tasks. To complete the work outlined in this assignment, you need to pay about $25. If paying $25 of your own money presents a financial hardship to you, then please email your professor.

We will be asking workers to label images for us, similar to how Fei-Fei Li created ImageNet.

Your first HIT

Let’s get started with creating your first HIT on MTurk. After logging into your requester account, go to the Create Tab and then click New Project. We’ll be labeling images, so you can start with the built-in Image Contains HIT design (shown below).

Click on the Create tab to start building a New Project

Select Image Classification, and then click on Create Project. You will see 3 tabs:

  • Enter Properties is where you can configure how much you will pay workers to do your task (Reward per assignment), how many workers you want to annotate each item (Number of assignments per task), and who is allowed to work on your your tasks (Worker requirements).
  • Design Layout is where you can edit the HTML that forms your HIT. You can update the instructions, add new html form elements to collect input from workers, and modify the variable names (denoted as ${variable_name}).
  • Preview and Finish is where you see how your task will look to workers. When you publish your tasks, any variables in the layout will be replaced with the input data that you provide via a comma separated value (CSV) file. You haven’t uploaded any data yet, so you won’t see it in the preview.

In the Enter Properties tab, you should change the following fields:

  • Project Name - this is a mneumonic name for yourself. I named mine “Image labeling version 1”.
  • Title - this is what the workers will see in the on the HIT Groups listing on worker.mturk.com. We’re going to be having workers label wedding photos, so I changed this field to say “Does this image show a wedding?”.
  • Description - this is what a worker sees once they click on your HIT title. I changed this field to say “Does this image depict a wedding? We are interested in both Indian weddings and Western-style weddings.”
  • Keywords - workers can search for tasks based on this. In addition to image, and classification, I added wedding.
  • Reward per assignment - this is how much you’ll pay a worker to do one task. It doesn’t include Amazon’s commission. I set this at $0.01.
  • Number of assignments per task - this is how many workers you want to label each item. I set this at 3 so that we can take a majortiy vote in cases where workers disagree about what the correct answer should be.

Edit the properties of your project

In the Worker requirements section of the Enter Properties tab, I added several qualifications that workers had to meet in order to do my task:

  • HIT Approval Rate >= 90%
  • Number of HITs approved >= 500
  • Location is India (since we’ll be labeling a lot of photos related to Indian weddings)

Edit the Worker Requirements for your project

In the Design Layout tab, edit the HTML to say that we’re looking for pictures of weddings instead of birds, and edit the Detailed Instructions section to give clearer instructions on what we’re looking for. Here’s what I wrote:

Instructions for determining if the image depicts a wedding:

  • We are interested in both Indian weddings and Western-style weddings
  • Pick yes if the image shows a bride and groom
  • Pick yes if the image shows a wedding or something related to a wedding ceremony

Edit the Worker Requirements for your project

After you’ve made those edits, you can click on Save and then Preview to see what your HIT will look like to workers. Since you haven’t yet uploaded any information (like the URLs for the images that you want users to judge), there will be a placeholder saying “Image will display here”.

Preview what your HIT will look like to workers

If you’re satisfied with how the HIT looks, you can click the Finish button. You’ll then see your newly created task listed with a big orange button saying “Publish Batch” next to it. You can publish a batch by clicking that button and uploading a comma separated value (CSV) file with the inputs to the HIT. The popup screen that appears will give you a link to download a sample .csv file that shows what fields you need for your HIT. For this HIT design, all we need is a single column with the header image_url, and a list of the image URLs that we want workers to judge. Here’s a small CSV file that you can use to test your HIT.

Upload your CSV file

Once you’ve uploaded the CSV then you can preview the HITs with the data populating them. There’s a “Next HIT” button that will let you click through and preview multiple assignments. This is useful to check things like whether your image links are all working properly.

Preview what your HIT will look like to workers

If you’re happy with how the HITs look, you can click the big orange “Next” button. You’ll then see a summary screen that gives details about the HIT including how much it will cost. You can then publish your task for Turkers to work on by clicking on the big orange “Publish” button.

Confirm the cost before you publish

Your HITs will be posted to MTurk. Once work has begun, you can mointor progress in the Manage tab. You’ll see a green progress bar showing how many of them have been completed.

Confirm the cost before you publish

You can see the individual responses by clicking on the “Review Results” link above the progress bar. On this screen you’ll see:

  • The HIT ID - notice that there are 3 HIT IDs that are all identical, since we asked 3 workers to annotate each item.
  • The Worker ID of the workers who completed each HIT.
  • The Approval Rate for the worker - this is how many of the worker’s tasks we have approved in the past (not how many all requesters have approved - just us).
  • Input.Image Url - the input field in the CSV that we uploaded.
  • Image Contains.Label - the label that the worker picked.

Review the results from the workers

You can see that the three workers all said that the first URL did not show a wedding. This is the image that they said doesn’t show a wedding:

This doesn't show a wedding

Three workers said that the second URL did show a wedding. This is the image that does show a wedding:

This does show a wedding

The Review Results screen will also let you approve or reject the Workers’ submissions. I recommend going ahead and approving all the assignments for this test HIT. You can also Download all of these results in a CSV file. Here is the results file from when I ran the HIT. You can open the CSV file in a spreadsheet app. You’ll notice that the CSV file has a lot of extra fields that include information about your HIT, including the properties that you included for it, and info about the amount of time that workers took to complete each assignment. Columns in the results CSV that start with “Input.” are the variables that were in the CSV that you uploaded. Columns that start with “Answer.” are the answers that the Turkers provided.

Why are we labeling images of Indian weddings?

Last week, you trained an image classifier with modern machine learning methods to achieve state-of-the-art results, making use of transfer learning from the ImageNet dataset, which was collected with large-scale crowdsourcing.

AI encodes and magnifies bias, and Google researchers found that ImageNet and another popular dataset called Open Images “appear to exhibit an observable amerocentric and eurocentric representation bias,” as demonstrated by the distribution of geographically identifiable images in the datasets, with 2/3 of the images from the Western world.

Chart from 'No Classification without Representation'

In addition, classifiers trained on the datasets show “strong differences in the relative performance on images from different locales”, with lower accuracy and confidence on images with labels related to people, like “bridegroom” and “police officer”, from countries like India and China. The research helped inspire the Inclusive Images Challenge, run by Google in partnership with a top deep learning conference called NeurIPS, last year.

This classifier has no problem correctly labeling photos of Western weddings, but fails on pictures of weddings from other cultures

A different large-scale crowdsourced dataset, The Massively Multilingual Image Dataset (MMID), was created by Penn researchers to learn English translations for words in 100 foreign languages, by scraping images for each foreign word and finding the English words that had the most “similar” images.

An image from the massively multilingual images dataset

MMID contains around 100 images for around 10,000 words in 100 foreign languages, providing an interesting source of data for improving the “geodiversity” of image classifiers. However, the images for an English translation of a foreign word can be noisy, as shown by crowdworkers who evaluated the relevance of images for a large subset of translations in 3 languages.

In this assignment, you will explore how a classifier pre-trained on ImageNet performs on photos representing wedding-related words in several Indian languages, and employ Indian workers on Mechanical Turk to validate that images are in fact related to weddings. You should work in teams of 2.

What to do in this assignment

You’ll do the following things in this assignment:

  1. Use an image classifier trained on ImageNet to classify a set of wedding pictures that were collected in the MMID data set for several Indian langauges and several Western European languages.
  2. See what faction of the images are predicted to depict weddings by the classifier.
  3. Post the images to Mechanical Turk and having India-based Turkers manually label whether the images depict weddings. We’ll give you a HIT design that will allow Turkers to label more than one image at once.
  4. Aggregate the workers results to create a set of correct labels for the images.
  5. Validate the model’s predictions by comparing them to the correct labels that you derived from the Turkers’ answers.
  6. Analyze the results and write a short report about your findings.
  7. Upload your MTurk results CSV file and your code.

Detailed Instructions

  1. In a Colab notebook with a GPU runtime (Runtime -> Change runtime type -> Hardware accelarator -> GPU), follow the Keras code to Classify ImageNet classes with ResNet50 on a wedding image you download from Google Images and upload to Colab, to get a feel for the code. Keras is a high-level neural networks library that makes it easy to run pre-trained models. You should upload files to Colab by going to the “Files” tab on the left panel.

  2. Upload the zipped “Weddings Indian Languages” dataset to Colab and run !unzip "weddings-indian-languages.zip" in a new cell. The dataset is composed of around 200-1000 images per language, for 8 languages spoken in India (Bengali, Gujarati, Hindi, Malayalam, Marathi, Punjabi, Tamil, and Telugu), taken from MMID. Repeat with the “Weddings European Language” dataset, which includes Spanish.

  3. Create a Pandas DataFrame from a list of dictionaries, where each dictionary contains the results of the classifier on an image, and looks like this.
    {"path": "weddings-indian-languages/punjabi/7899/01.jpg",
    "predictions": ["vestment", "abaya", "cloak"],
    "predictions_include_groom_or_bridegroom": False}
    

    We recommend using the glob module with the appropriate wildcards to get a list of all the images. Save the DataFrame as image_paths_and_predictions.csv, which you will use later in the assignment. To simplify step 7, you must add “https://s3.amazonaws.com/nets213-hw5/” to the beginning of each image file path, before saving the DataFrame as a CSV.

  4. We have created an alternate HIT design that lets workers label 12 images at a time (here’s a screenshot). Use the HIT design that we provide to create the HIT. Download a sample of the input CSV file for the project at the top of the preview page, and finish creating the HIT.

  5. Use the sample input.csv file format and data from image_paths_and_predictions.csv (created in step 3, which you can load in as a DataFrame) to create variables.csv with the right format for this HIT.

  6. Click “Publish Batch” in MTurk, uploading variables.csv, and preview the tasks. Click “Next” and confirm the settings of your HIT, which should cost approximately $25 per team. Make sure to screenshot this page for the report. Sit back and watch the crowd work!

  7. When the HIT is done, download the Batch CSV and read it into a DataFrame in Colab. For every row in the DataFrame, split “Answer.selected” to get the list of images that workers identified as wedding-related. For each image in the row (in columns “Input.image"), if image is in the selected images, update a counter, where the key is the URL in the "Input.image" column. Here is the pseudocode:
    Create a Counter object counts
    For every row in the DataFrame:
      true_images = the list of images from splitting the string in the "Answer.selected" column of the row
      for every column "Input.image<number>":
          url = row["Input.image<number>"]
          if image<number> is in true_images:
             counts[url] += 1
          else:
             counts[url] += 0
    
  8. Create a DataFrame from the resulting counter, and derive a new column that is True only if the counter value is 2 or more (a majority of the workers said the image represented weddings). Use the merge function to join the DataFrame loaded from image_paths_and_predictions.csv to the DataFrame of true labels, on the column of image paths. Save the DataFrame as submissions.csv. Calculate the precision, recall, and F1 score of the classifier, for Western images vs. non-Western images. Are you surprised by the results you got? Analyze the predictions and “true labels” further by visualizing images that the classifier (in)correctly labeled.

Report

Below are the questions that you will be asked to answer about this assignment. Please turn in your answers in a PDF for Homework 5 on Gradescope.

  1. What is the link to your Colab notebook?
  2. Attach a screenshot of the page confirming the settings of your HIT.
  3. What is the precision, recall, and F1 score of the classifier on Western images vs. on non-Western images? Compare the differences.
  4. Analyze how the predictions of the classifier compare to the labels of the workers. Include images to explain why you think the classifier correctly or incorrectly labeled certain images.
  5. If you had more time to work on this HIT, what additional things would you add in the creation or processing of the HIT for better quality control and aggregation?
  6. Upload submissions.csv and the Batch CSV file (From step 9) with the MTurk results
  7. Calculate the cost of annotating images from the entire wedding dataset (Both Indian and European) on MTurk using the 1-image-per-HIT design that was described in the section “Your First HIT”. Compare this cost with the cost you obtained while using the multiple-images-per-HIT-design we gave you. Assume the “Reward per Assignment” in both the cases to be $0.01, and the “Number of Assignments per Task” to be 3. State any other assumptions necessary.
  8. Calculate the cost of annotating the 35 million images in the MMID. For this question, you are free to choose any HIT design, and any values for “Reward per Assignment” and “Number of Assignments per task”. State your reasons for choosing the design and these values. State any other assumptions necessary.