Car Retrieval

Baseline codes and explanations for the car retrieval task.

Task explanation

Information retrieval is a task for searching images containing the same information of the input image. You have a "query" image (input image) and "gallery" or "candidate" images (images in the database).

Consider an example in the below table:

	input	sample1	sample2
samples
classes	class_A	class_A	class_B

Your model should distinguish that input and sample1 are in the same class (class_A), while sample2 is in a different class (class_B).

We assume that your model doesn't know any information in the test set, i.e., your model shouldn't be a simple classifier. We provide a simple baseline retrieval model based on ResNet-18.

Evaluation explanation

In this task, your model will return the nearest sample of the given query image. We will measure the top-1 classification accuracy.

Consider you have four images with class [dog, dog, cat, cat] and assume your model predict the nearest sample for every image as the following.

query image id	the nearest image id	query class	nearest class	correct?
Image1	Image2	dog	dog	Yes
Image2	Image3	dog	cat	No
Image3	Image4	cat	cat	Yes
Image4	Image3	cat	cat	Yes

In this case, your model has a score 0.75 = 3/4.

For the baseline score, our provided baseline model achieves 0.8449212384573601.

Dataset explanation

Split	Number of Classes	Number of Samples
Train	84	53,813
Test	78	11,046
Test submit	4	20

• Image resolution: 512 x 384
• Environments
  • Rotations (15 degree interval. [0, 15, ..., 345])
  • Time zone (morning, noon, afternoon, and night)

In this task, we set the "name of car" as classes. For example, "아반떼 HD (red)" and "아반떼 MD (silver)" are in the same class "아반떼."

Step-by-step tutorial

Run your code

nsml run  -d 9_iret_car

More tips

• You can specify more options for nsml! See nsml run --help.
• You can check your progress using nsml logs [your_session_name]
  • e.g., nsml logs username/9_iret_car/1

Prepare submit

nsml model ls [your_session_name]
# e.g., nsml model ls username/9_iret_car/1

Check your models.

Submit your model

nsml submit  [your_session_name] [model_name]
# e.g., nsml model ls username/9_iret_car/1 100

It will take about 5 minutes.

Baseline model

Here is detailed explanation of our example model.

• We train ResNet-18 using classification loss on the training set.
• We extract the features from the last layer of trained ResNet-18.
• We select the nearest sample using the cosine similarity between a query image and candidate images.