A Unified Multimodal Benchmark

Design

Abilities

This benchmark evaluate the MLLMs's abilities from 3 levels:

• Corse-grained Conceptually Understanding

  Know the world basically: align vision to concepts

  • recognize objects with conceptual classes and identify their spatial relationships
  • recognize the positions of objects
  • counting objects with conceptual classes
  • describing the existence and activity in visual scene
  • recognize composition of objects

• Fine-grained Specifically Understanding**

  Know the world specifically: align vision to entities

  • align objects to world entities
  • align composition of objects to world entities
  • align relationships between objects with semantic relations
  • align visual scene with world actual events

• Associational Understanding

  Know the world thoroughly: image, reasoning and planing on current visual scene

  • associate objects with other similar objects that do not exist in current image
  • associate relations with other similar relations
  • associate composition of objects with other compositions
  • associate events with other similar events
  • metaphor, joke
  • hallucination
    • knowledge hallucination
    • existence hallucination
    • attribute hallucination
  • Interactions in Embodied Environment

Code

For the sake of readability, some details have been omitted.

.
├── README.md
├── data                    # data processor, raw data -> processed data
└── mmbench
    ├── common
    │   ├── example.py
    │   ├── training.py     # main training, ref sat
    │   ├── inference.py    # main testing, ref sat
    │   ├── model.py        # model interface
    │   ├── registry.py     # registry
    │   └── utils.py        # common functions
    ├── metrics             # all metrics
    │   ├── bleu
    │   └── rouge
    │   └── acc
    │   └── vqa_acc
    └── tasks               # all tasks
    │   ├── base_task.py    # main task, including most of functions in evaluating
    │   ├── level_1
    │   │   ├── VQAv2
    │   │   └── Visual7W
    │   │   └── ...
    │   ├── level_2
    │   │   └── OK-VQA
    │   └── level_3
    │       └── HalVQA
    ├── __init__.py
    ├── config.yaml         # configure data paths, params, and other
    └── evaluator.py        # main entry

Usage

Install

Install this repo from source.

git clone git@github.com:qijimrc/mm_evaluation.git
cd mm_evaluation & python3 setup.py install

Example

import argsparse
from mmbench.evaluator import Evaluator
from mmbench.common.model import ModelInterface

parser = argparse.ArgumentParser()
parser.add_argument('--eval_tasks', type=str, nargs='+', help='Specify the tasks for evaluation')
parser.add_argument("--custom_cfg_path", type=str, help="customized eval config path")
args = parser.parse_args()

# build your ModelInterface
mt = ModelInterface(args, model, ...)
# Evaluate
evaluator = Evaluator(custom_cfg_path=args.custom_cfg_path, custom_functions={})
scores = evaluator.evaluate(args, mt, eval_tasks=args.eval_tasks)

Features

1. customized params

Create a customized yaml config referring tasks in the mmbench/config.yaml. Then add the custom_cfg_path in the args when you build the Evaluator.

2. customized functions

You can customized the following functions in the mmbench/tasks/base_task.py

• preprocess_datab_eval
• collate_fn
• forward_step
• forward_step_eval

Leaderboard

Model	level_1			level_2			level_3			AVG
	VQAv2						HalVQA
BLIP2
LLaVA
VisualGLM-6B