GeoNet is an end-to-end unsupervised learning framework for estimating monocular depth, optical flow, and camera motion from video sequences. The framework leverages the geometric relationships between these tasks, learning them jointly through a deep convolutional network. The model uses a divide-and-conquer strategy, with a two-stage architecture: a rigid structure reconstructor for static scene geometry and a non-rigid motion localizer for dynamic objects. A novel adaptive geometric consistency loss is introduced to enhance robustness against outliers and non-Lambertian regions, effectively resolving occlusions and texture ambiguities. The model is evaluated on the KITTI dataset, achieving state-of-the-art results in all three tasks, outperforming previous unsupervised methods and performing comparably with supervised ones. The rigid structure reconstructor uses depth maps and camera poses to estimate rigid flow, while the non-rigid motion localizer handles dynamic objects with a residual flow learning module. The geometric consistency enforcement ensures that predictions are coherent across different views, improving the model's performance in occluded and texture ambiguous regions. The model's unsupervised nature allows it to learn from raw video data without requiring ground truth, making it efficient and scalable. Experiments show that GeoNet outperforms other methods in monocular depth estimation, optical flow estimation, and camera pose estimation. It achieves lower end-point error (EPE) in overall regions and comparable results in non-occluded regions. The model's ability to handle dynamic objects and occlusions is demonstrated through visual comparisons. Additionally, GeoNet outperforms traditional SLAM frameworks in camera pose estimation, showing its effectiveness in capturing high-level cues from 3D scene geometry. The framework's unsupervised learning approach allows it to learn from raw video data without requiring ground truth, making it efficient and scalable. The model's ability to handle dynamic objects and occlusions is demonstrated through visual comparisons. The results show that GeoNet is effective in capturing high-level cues from 3D scene geometry, outperforming other methods in various tasks. The model's architecture and training strategy enable it to achieve state-of-the-art results in monocular depth, optical flow, and camera pose estimation.GeoNet is an end-to-end unsupervised learning framework for estimating monocular depth, optical flow, and camera motion from video sequences. The framework leverages the geometric relationships between these tasks, learning them jointly through a deep convolutional network. The model uses a divide-and-conquer strategy, with a two-stage architecture: a rigid structure reconstructor for static scene geometry and a non-rigid motion localizer for dynamic objects. A novel adaptive geometric consistency loss is introduced to enhance robustness against outliers and non-Lambertian regions, effectively resolving occlusions and texture ambiguities. The model is evaluated on the KITTI dataset, achieving state-of-the-art results in all three tasks, outperforming previous unsupervised methods and performing comparably with supervised ones. The rigid structure reconstructor uses depth maps and camera poses to estimate rigid flow, while the non-rigid motion localizer handles dynamic objects with a residual flow learning module. The geometric consistency enforcement ensures that predictions are coherent across different views, improving the model's performance in occluded and texture ambiguous regions. The model's unsupervised nature allows it to learn from raw video data without requiring ground truth, making it efficient and scalable. Experiments show that GeoNet outperforms other methods in monocular depth estimation, optical flow estimation, and camera pose estimation. It achieves lower end-point error (EPE) in overall regions and comparable results in non-occluded regions. The model's ability to handle dynamic objects and occlusions is demonstrated through visual comparisons. Additionally, GeoNet outperforms traditional SLAM frameworks in camera pose estimation, showing its effectiveness in capturing high-level cues from 3D scene geometry. The framework's unsupervised learning approach allows it to learn from raw video data without requiring ground truth, making it efficient and scalable. The model's ability to handle dynamic objects and occlusions is demonstrated through visual comparisons. The results show that GeoNet is effective in capturing high-level cues from 3D scene geometry, outperforming other methods in various tasks. The model's architecture and training strategy enable it to achieve state-of-the-art results in monocular depth, optical flow, and camera pose estimation.