In certain situations, images need to be interpreted within a broader context. We introduce a new benchmark, named as CODIS (COntext-Dependent Image diSambiguation), designed to assess the ability of models to use context provided in free-form text to enhance visual comprehension. It stands out from existing benchmarks in three main aspects:

We introduce a new benchmark, named as CODIS (COntext-Dependent Image diSambiguation), designed to assess the ability of models to use context provided in free-form text to enhance visual comprehension. We have identified five representative types of context, including three types of global context that pertain to the overall scene—namely, the global background, which encompasses location and orientation, temporal information, and cultural background, and two types of local context related to objects within the scene, specifically the attributes of objects and the relationships between people.

To prevent models from guessing the correct answers without fully understanding context, we organize our dataset in pairs. Each pair contains two queries $ (I,Q,C_1) $ and $ (I,Q,C_2) $. The queries have identical image $ I $ and question $ Q $, but have two pieces of different context $ C_1 $ and $ C_2 $. We give MLLMs two queries separately, and get model outputs $ O_1 $ and $ O_2 $.

We leverage two metrics, pair-wise accuracy $ \mathrm{Acc}_p $ and query-wise accuracy $ \mathrm{Acc}_q $ for our evaluation metrics. For $ \mathrm{Acc}_p $, models score only if their answers to a pair of queries are both correct. For $ \mathrm{Acc}_q $, models score for each single query they answer correctly.

We manually collect images that contain ambiguities which can only be resolved with external contexts. The majority of these images are real-scene images from the publicly available dataset ShareGPT4V and the Internet, while the remainder are created manually. For each collected image, we manually write questions, context and answers for it.

We summarize recent benchmarks for MLLMs in the following table. Most of these benchmarks do not pair images with additional context. Only two benchmarks, namely VisDial and MMDialog, include extra context to help in conversation with humans rather than to clarify the meaning of images. This limitation means these benchmarks are not fully capable of testing the ability of MLLMs to understand images in a context-dependent manner.

We collect 216 images and successfully collected a total of 706 queries, spanning five categories and encompassing a wide range of scenarios. The distribution of categories and scenarios is illustrated in the following figure.