CoDraw: Collaborative Drawing as a Testbed for Grounded Goal-driven Communication

Abstract

In this work, we propose a goal-driven collaborative task that combines language, perception, and action. Specifically, we develop a Collaborative image-Drawing game between two agents, called CoDraw. Our game is grounded in a virtual world that contains movable clip art objects. The game involves two players: a Teller and a Drawer. The Teller sees an abstract scene containing multiple clip art pieces in a semantically meaningful configuration, while the Drawer tries to reconstruct the scene on an empty canvas using available clip art pieces. The two players communicate with each other using natural language. We collect the CoDraw dataset of ∼10K dialogs consisting of ∼138K messages exchanged between human players. We define protocols and metrics to evaluate learned agents in this testbed, highlighting the need for a novel crosstalk evaluation condition which pairs agents trained independently on disjoint subsets of the training data. We present models for our task and benchmark them using both fully automated evaluation and by having them play the game live with humans.

Multimodal Intelligence: Representation Learning, Information Fusion, and Applications

“…It is shown with minor modifications, Text2Scene can generate cartoon like, semantic layout, and real image like scenes. Dialogue based interaction is studied to control image synthesis, in order to improve complex scene generation progressively [219]- [223]. Meanwhile, text-to-image synthesis is extended to multiple images or videos, where visual consistency is required among the generated images [224]- [226].…”

Tell, Draw, and Repeat: Generating and Modifying Images Based on Continual Linguistic Instruction

“…For this task, we use the synthetic Collaborative Drawing (CoDraw) dataset [8], which is composed of sequences of images along with associated dialogue of instructions and linguistic feedback ( Figure 2). Also, we introduce the Iterative CLEVR (i-CLEVR) dataset (Figure 4), a modified version of the Compositional Language and Elementary Visual Reasoning (CLEVR) [9] dataset, for incremental construction of CLEVR scenes based on linguistic instructions.…”

“…The most similar task to GeNeVA is the task proposed by the CoDraw [8] authors. They require a model to build a scene by placing the clip art images of the individual objects in their correct positions.…”

“…CoDraw [8] is a recently released clip art-like dataset. It consists of scenes, which are sequences of images of children playing in a park.…”

“…boys hand covers trunk. An example from the CoDraw[8] dataset. Each example from the dataset involves a conversation between a Teller and a Drawer.…”