Image Segmentation
Transformers
PyTorch
pixdlm
cvpr-2026
compute-transparency
reasoning-segmentation
uav
remote-sensing
vision-language
Instructions to use WhynotHug/PixDLM with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use WhynotHug/PixDLM with Transformers:
# Use a pipeline as a high-level helper from transformers import pipeline pipe = pipeline("image-segmentation", model="WhynotHug/PixDLM")# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("WhynotHug/PixDLM", device_map="auto") - Notebooks
- Google Colab
- Kaggle
Upload README.md
Browse files
README.md
CHANGED
|
@@ -22,12 +22,7 @@ tags:
|
|
| 22 |
[](https://huggingface.co/datasets/WhynotHug/DRSeg)
|
| 23 |
[](https://huggingface.co/spaces/WhynotHug/PixDLM)
|
| 24 |
|
| 25 |
-
PixDLM is a dual-path multimodal language model for UAV reasoning segmentation.
|
| 26 |
-
Given an aerial image and a reasoning-oriented question, PixDLM produces a target
|
| 27 |
-
segmentation mask aligned with the described object.
|
| 28 |
-
|
| 29 |
-
This repository contains the clean open-source release for the CVPR 2026
|
| 30 |
-
Highlight paper and the 2027 CVPR Compute Transparency Champion release.
|
| 31 |
|
| 32 |
## Highlights
|
| 33 |
|
|
@@ -160,28 +155,6 @@ logs/<exp_name>/
|
|
| 160 |
Typical per-sample artifacts include the input image, predicted mask, ground
|
| 161 |
truth mask, red/green overlay, and a JSON result file.
|
| 162 |
|
| 163 |
-
## DRSeg Evaluation Results
|
| 164 |
-
|
| 165 |
-
The paper reports the following PixDLM metrics on DRSeg:
|
| 166 |
-
|
| 167 |
-
| Reasoning type | gIoU | cIoU |
|
| 168 |
-
| --- | ---: | ---: |
|
| 169 |
-
| Attribute | 62.80 | 62.84 |
|
| 170 |
-
| Scene | 61.75 | 64.03 |
|
| 171 |
-
| Spatial | 62.51 | 62.80 |
|
| 172 |
-
|
| 173 |
-
A practical 8-GPU verification run with this release produced:
|
| 174 |
-
|
| 175 |
-
| Reasoning type | gIoU | cIoU |
|
| 176 |
-
| --- | ---: | ---: |
|
| 177 |
-
| Attribute | 62.51 | 60.82 |
|
| 178 |
-
| Scene | 62.56 | 62.36 |
|
| 179 |
-
| Spatial | 61.54 | 61.00 |
|
| 180 |
-
| Overall | 62.20 | 61.41 |
|
| 181 |
-
|
| 182 |
-
The gIoU reproduction is close to the paper table. The cIoU numbers can vary with
|
| 183 |
-
the exact evaluation protocol, sampler behavior, and dependency versions.
|
| 184 |
-
|
| 185 |
## Dataset
|
| 186 |
|
| 187 |
DRSeg is available at:
|
|
|
|
| 22 |
[](https://huggingface.co/datasets/WhynotHug/DRSeg)
|
| 23 |
[](https://huggingface.co/spaces/WhynotHug/PixDLM)
|
| 24 |
|
| 25 |
+
【CVPR2026 Highlight】PixDLM is a dual-path multimodal language model for UAV reasoning segmentation.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
|
| 27 |
## Highlights
|
| 28 |
|
|
|
|
| 155 |
Typical per-sample artifacts include the input image, predicted mask, ground
|
| 156 |
truth mask, red/green overlay, and a JSON result file.
|
| 157 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 158 |
## Dataset
|
| 159 |
|
| 160 |
DRSeg is available at:
|