Unverified50% confidenceSolutionExact time
The author selected 500 frames of images to manually correct annotation boxes and mapped RGB annotations to infrared images by leveraging the coaxial calibration property between LWIR and RGB.
1
Sources
50%
Confidence
Long-term
Relevance
8/9/2026
First Seen
Sources
自建多模态数据集：从500帧到300万帧的自动化标注实践
Related Claims
UnverifiedThe author's V1 model completed pre-training on RGB images using the COCO dataset, with 500 frames of images selected for manual correction of annotation boxes.77% similarUnverifiedBy leveraging the bore-sighted alignment between LWIR and RGB, annotations on RGB can be directly mapped to infrared images, achieving zero-cost transfer of a single annotation to another modality.75% similarUnverified亚像素渲染通过对RGB子像素的独立控制，可以在视觉上实现比物理像素更细腻的边缘过渡67% similarUnverifiedVersatIL supports pluggable encoders for RGB images, depth maps, proprioception, and language observations.65% similarUnverifiedWhen a reference image is uploaded, the model encodes it into a high-dimensional feature vector carrying visual information about style, tone, and composition.61% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/714109API
curl https://kongchang.com/api/v1/knowledge/claims/714109MCP
get_claim(id=714109)