Research map / Vision-language and multimodal / Multimodal LLMs

PixDLM: A Dual-Path Multimodal Language Model for UAV Reasoning Segmentation

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Spatial reasoning in VLMs · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Defines UAV reasoning segmentation, releases the DRSeg benchmark (10k UAV images with CoT QA and masks), and proposes PixDLM, a dual-path (CLIP+SAM) LISA-style pixel-level MLLM baseline.

Problem

Reasoning segmentation under UAV conditions (oblique views, ultra-high resolution, extreme scale variation) lacks a dedicated benchmark and baselines, since existing models rely on low-resolution tokens and ground/satellite data.

Approach

A dual-path vision encoder (CLIP global path plus SAM high-resolution path) with gated three-stage latent fusion (SAM features injected into CLIP stages) and a hierarchical mask decoder, driven by a LLM mask token, plus a new CoT-annotated UAV reasoning segmentation benchmark DRSeg.

Key results

• PixDLM with SFT reaches 62.80/62.84 gIoU/cIoU on Attribute, 61.75/64.03 on Scene, and 62.51/62.80 on Spatial reasoning in DRSeg, above SFT LISA-7B and PixelLM-7B.
• Zero-shot baselines such as LISA-7B reach only about 42-53 gIoU on DRSeg subsets, and SegEarth-R1 reaches 20-26 gIoU.
• Human verification on 2,000 images reports 95% uniquely identifiable questions, CoT score 2.83/3, and 97% mask alignment.

Surprising findings

• Adding RRSIS-D to DRSeg training gives limited gains, which the authors attribute to weaker reasoning annotations.
• Corrupting, masking, or shuffling CoT traces on 10% of samples causes only limited performance change, suggesting robustness to noisy reasoning supervision.

Limitations

• Each image is annotated with only one target instance, so multi-target and dense-object scenarios are not covered by the reasoning labels.
• Reasoning annotations are generated by GPT-5 with human verification on a sample, so annotation bias is possible.
• Zero-shot results for several baselines are low, and the paper gives limited analysis of failure cases.

Experimental setup

Datasets: DRSeg RRSIS-D CODrone RefCOCO RefCOCO+ RefCOCOg

Benchmarks: DRSeg RefCOCO RefCOCO+ RefCOCOg

Metrics: gIoU cIoU

Baselines: SegEarth-R1 LISA-7B LISA-13B-llama2-v1 PixelLM-7B PixelLM-13B LISAt-7B Seg-Zero-7B GeoPix-7B GSVA-7B LaSagnA-7B

Builds on: LISA PixelLM LLaVA-HR SAM2 CODrone

Closest papers by meaning

• DR$^2$Seg: Decomposed Two-Stage Rollouts for Efficient Reasoning Segmentation in Multimodal Large Language Models (ICML 2026 · similarity 0.86)
• PixelLM: Pixel Reasoning with Large Multimodal Model (CVPR 2024 · similarity 0.86)
• LISAt: Language-Instructed Segmentation Assistant for Satellite Imagery (NeurIPS 2025 · similarity 0.85)
• Unleashing Multispectral Video's Potential in Semantic Segmentation: A Semi-supervised Viewpoint and New UAV-View Benchmark (NeurIPS 2024 · similarity 0.85)
• RPTS: Tree-Structured Reasoning Process Scoring for Faithful Multimodal Evaluation (AAAI 2026 · similarity 0.85)
• High Temporal Consistency through Semantic Similarity Propagation in Semi-Supervised Video Semantic Segmentation for Autonomous Flight (CVPR 2025 · similarity 0.85)