See, Say, and Segment: Teaching LMMs to Overcome False Premises

TL;DR

This work tackles false-premise errors in open-domain reasoning segmentation by introducing FP-RefCOCO and the SESAME model, which can See, Say, and Segment via a cascaded and jointly trained LMM framework. FP-RefCOCO provides context-aware negative prompts and corrected alternatives to train robust false-premise handling, while SESAME integrates object detection, corrective dialogue, and segmentation without catastrophic forgetting. Empirical results show substantial gains in false-premise detection, natural-language corrections, and segmentation accuracy, especially under high false-premise prevalence, and the method maintains strong performance on traditional positive-only benchmarks. The work advances interactive, grounded language-vision systems with practical implications for robust visual reasoning in real-world user interactions.

Abstract

Current open-source Large Multimodal Models (LMMs) excel at tasks such as open-vocabulary language grounding and segmentation but can suffer under false premises when queries imply the existence of something that is not actually present in the image. We observe that existing methods that fine-tune an LMM to segment images significantly degrade their ability to reliably determine ("see") if an object is present and to interact naturally with humans ("say"), a form of catastrophic forgetting. In this work, we propose a cascading and joint training approach for LMMs to solve this task, avoiding catastrophic forgetting of previous skills. Our resulting model can "see" by detecting whether objects are present in an image, "say" by telling the user if they are not, proposing alternative queries or correcting semantic errors in the query, and finally "segment" by outputting the mask of the desired objects if they exist. Additionally, we introduce a novel False Premise Correction benchmark dataset, an extension of existing RefCOCO(+/g) referring segmentation datasets (which we call FP-RefCOCO(+/g)). The results show that our method not only detects false premises up to 55% better than existing approaches, but under false premise conditions produces relative cIOU improvements of more than 31% over baselines, and produces natural language feedback judged helpful up to 67% of the time.

Figures (9)

• Figure 1: False premise failures with LMMs: contemporary open-source LMMs combined with segmentation decoders are able to generate referring segments effectively but have difficulty on segmentation questions which ask the model to refer to something that is not present in the image. SESAME, our See-Say-Segment LMM, uses model chaining and joint training to overcome this problem.
• Figure 2: SESAME is an LMM that can "see" whether objects are detected in an image and "say" by telling the user if they are there or not. When appropriate, alternative queries can be offered or semantic errors corrected in the query. SEASAME can then "segment" by returning the mask of the desired object.
• Figure 3: FP-RefCOCO Dataset Creation. Using refCOCO for base images, we employ an LLM to create a false-premise referring segmentation dataset with similar objects, attributes, and relations. Such paired examples enable the the creation of specific correction ground truth that is more specific than baseline methods which simply sample positive and negative examples. This data allows us to train an LMM that has robust reasoning reference capabilities.
• Figure 4: In contrast to prior work (the output of the LISA lai2023lisa is shown above), SESAME is able to handle more complicated conditional reasoning and instruction, and is able to not output a segment when it is not requested.
• Figure 5: Not only is SESAME robust to false premises, and does not attempt to incorrectly predict a segmentation mask when an object or concept is not actually present in the image, but it is able to use commonsense reasoning to suggest relevant objects or concepts when a similar instance is present, and output the segmentation mask of that instance.