Benchmark that tests shape recognition


Keywords
tests, shape, recognition, capacity
License
MIT
Install
pip install ShapeY==0.1.8

Documentation

ShapeY

ShapeY is a benchmark that tests a vision system's shape recognition capacity. ShapeY currently consists of ~68k images of 200 3D objects taken from ShapeNet. Note that this benchmark is not meant to be used as a training dataset, but rather serves to validate that the visual object recogntion / classification under inspection has developed a capacity to perform well on our benchmarking tasks, which are designed to be hard if the system does not understand shape.

Installing ShapeY

Requirements: Python 3, Cuda version 10.2 (prerequisite for cupy)

To install ShapeY, run the following command:

pip3 install shapey==0.1.7

Step0: Download ShapeY200 dataset

Run download.sh to download the dataset. The script automatically unzips the images under data/ShapeY200/. Downloading uses gdown, which is google drive command line tool. If it does not work, please just follow the two links down below to download the ShapeY200 / ShapeY200CR datasets.

ShapeY200: https://drive.google.com/uc?id=1arDu0c9hYLHVMiB52j_a-e0gVnyQfuQV

ShapeY200CR: https://drive.google.com/uc?id=1WXpNUVRn6D0F9T3IHruml2DcDCFRsix-

After downloading the two datasets, move each of them to the data/ directory. For example, all of the images for ShapeY200 should be under data/ShapeY200/dataset/.

Step1: Extract the embedding vectors from your own vision model using our dataset

Implement the function your_feature_output_code in step1_save_feature/your_feature_extraction_code.py. The function should take in the path to the dataset as input and return two lists - one for the image names and another for the corresponding embedding vectors taken from whatever system.

Step2: Run macro.py

After you have implemented the function, run macro.py to generate the results. macro.py will automatically run the following steps:

  1. Compute correlation between all embedding vectors (using step2_compute_feature_correlation/compute_correlation.py)

  2. Run benchmark analysis (using step3_benchmark_analysis/get_nn_classification_error_with_exclusion_distance.py)

  3. Graph results (top1 object matching error, top1 category matching error, etc.)