 |
MIT CSAIL6.819/6.869: Advances in Computer Vision |
 |
Fall 2017 |
||
The goal of this challenge is to identify the scene category depicted in a photograph. The data for this task comes from the Places2 dataset which contains 10+ million images belonging to 400+ unique scene categories. Specifically, the mini challenge data for this course will be a subsample of the above data, consisting of 100,000 images for training, 10,000 images for validation and 10,000 images for testing coming from 100 scene categories. The images will be resized to 128*128 to make the data more manageable. Further, while the end goal is scene recognition, a subset of the data will contain object labels that might be helpful to build better models.
For each image, algorithms will produce a list of at most 5 scene categories in descending order of confidence. The quality of a labeling will be evaluated based on the label that best matches the ground truth label for the image. The idea is to allow an algorithm to identify multiple scene categories in an image given that many environments have multi-labels (e.g. a bar can also be a restaurant) and that humans often describe a place using different words (e.g. forest path, forest, woods). The exact details of the evaluation are available on the Places2 challenge website.
Students should improve the classification accuracy of their network models on the validation set of mini places challenge. The evaluation server is online already, so that students could submit their predictions of the test set for evaluations and ranking in the challenge leaderboard (see below for submission instructions).
We encourage students to use Amazon's EC2 for computation if they do not have access to their own GPUs to train deep networks. Students can sign up to receive free $100 credit through the AWS Educate program. We encourage students to use g2.2xlarge instances running Ubuntu for maximal ease of installing. Note that $100 of Amazon credit allows you to run a g2.2xlarge GPU instance for approximately 6 days without interruption (you should keep it on only while using it). In a larger group, you will get more total available compute time.
Also See
For more details, see https://github.com/CSAILVision/miniplaces.Please Use Only the Data Provided Here
For this course challenge, the dataset is different from the Places2 Challenge dataset. You do not need to register for the Places2 Challenge or download that data. You can only use the data provided in this challenge to train your models. You cannot use models that have been pre-trained using other datasets e.g., ImageNet or the full Places database.
Team Registration & Leaderboard
- A team of up to two students need to sign up with a team name here. The team members will receive a "teamcode"---token with which you submit your results---and instructions for submitting to the leaderboard. Working individually is also allowed, but it will be graded by the same standards.
- Submit your results to here. Only one submission is allowed for each team every 4 hours.
- Here is the leaderboard.
Grading Policy & Submission Instructions
- Due on Nov. 21, 11:59 pm. We strongly encourage you to start early, as you may need some time to set up the environment and familiarize yourself with deep learning tools.
- This challenge weighs two psets. You could drop this challenge, but that will use up your "drop quota" of two psets. We strongly recommend you to complete this challenge, as it will expose you to deep learning and get you ready for a deep learning final project.
- Each team can have up to two students. Feel free to look for teammates on Piazza.
- One team submits (1) one .pdf report with filename being "kerberos1_kerberos2_teamname.pdf" and (2) a .zip file of code named "kerberos1_kerberos2_teamname.zip". Only one person needs to submit. Please also include the kerberoses of the two team members and the teamname in the report.
- 6.819 and 6.869 students can mix up and form a team; their submission will be evaluated by 6.869 standards. Please specify 6.819 or 6.869 for each team member in the report.
Warning: Given the class size, we write scripts to crawl down your submissions, so if you fail to follow these guidelines (e.g., wrong file names, only one .zip submitted, .zip containing data and therefore gigantic in size), our script will skip your submissions, in which case you will be heavily penalized.
Data & Starter Code
Data (460MB)Development Kit (Github)
TensorFlow MiniPlaces sample training code
MatConvNet MiniPlaces reference network
Torch Starter Package
Tips
- It is very helpful to go through the examples of training convolutional neural nets.
- Use data augmentation, deeper layers, or object annotation etc, to boost the classification accuracy.