CNOCS Map: 3D Pose & Control Encoding

• CNOCS maps are dense, per-pixel geometric encodings that represent an object's 9-DoF pose (translation, scale, rotation) using normalized coordinates in a cuboid framework.
• They are computed by intersecting camera-space rays with a cuboid, transforming into the object’s local frame, and normalizing coordinates, ensuring a unique and invertible mapping.
• Integrated as conditional control signals in neural networks, CNOCS maps enhance spatial localization and pose accuracy for multi-object, multi-DoF scenarios.

A CNOCS map is a dense, per-pixel geometric encoding used for 3D object pose representation and neural conditional control in vision and generative models. The term stands for Cuboid Normalized Object Coordinate System, and the map unambiguously encodes a 9-degree-of-freedom (DoF) pose of a cuboid object—encompassing translation, scale, and rotation—across an image in a manner directly usable for model conditioning. Unlike earlier pose embeddings or NOCS (Normalized Object Coordinate Space) maps, CNOCS is optimized for category-agnostic, multi-object, and multi-DoF scenarios, offering explicit geometric interpretability and strong performance advantages in controllable vision generation tasks (Qin et al., 20 Nov 2025).

1. Definition and Semantic Structure

A CNOCS map for an object is a multi-channel image (typically 3 channels) where each foreground pixel encodes the normalized coordinates $(x, y, z)$ of the point where the corresponding camera ray intersects the object’s canonical unit-cuboid, given a specified 9-DoF pose. The channels correspond to the following:

Channel	Value at pixel $(u,v)$	Geometric meaning
1	Normalized $x$ in local cuboid	Canonical cuboid $x$-axis
2	Normalized $y$	Canonical cuboid $y$-axis
3	Normalized $z$	Canonical cuboid $z$-axis

For each object, one creates a separate CNOCS map, and in multi-object settings, maps are stacked depth-sorted channel-wise for network input (Qin et al., 20 Nov 2025).

2. Mathematical Formulation

Given a pose vector $$p = (t_x, t_y, t_z,\, s_x, s_y, s_z,\, r_{xy}, r_{yz}, r_{zx}) \in \mathbb{R}^9$$, with

• $l=(t_x, t_y, t_z)$ as object center in camera coordinates,
• $(u,v)$0 as cuboid side lengths,
• $(u,v)$1 as orientation (Euler angles),

the CNOCS mapping $(u,v)$2 is defined by:

1. For each pixel $(u,v)$3, compute the camera-space ray $(u,v)$4, with $(u,v)$5 the camera intrinsics.
2. Intersect $(u,v)$6 with the cuboid (centered at $(u,v)$7, scaled and rotated by $(u,v)$8); let $(u,v)$9 be the intersection.
3. Transform into the object local frame: $x$0.
4. Normalize per-dimension: $x$1 (and similarly for $x$2).
5. Assign: $x$3 for valid intersections, or $x$4 for background.

This process produces a unique, invertible mapping from the object pose to a dense, image-space geometric code (Qin et al., 20 Nov 2025).

3. Geometric Interpretation and Uniqueness

A CNOCS map provides a per-pixel bijection between the image plane and canonical object coordinates. Each pixel’s value represents the normalized $x$5 location in a cuboid aligned with the object’s specified pose. This construction encodes all pose DOFs:

• Translation: Controls which image region is “occupied” by the map.
• Scale: Affects the spatial span and normalization range.
• Rotation: Alters the canonical coordinate pattern in the map.

No two different poses yield the same CNOCS map. This property is critical for precise and disentangled pose control in structured neural inference and generative processes (Qin et al., 20 Nov 2025).

4. Network Integration and Training Regimes

CNOCS maps are incorporated as conditional control signals in a multi-branch neural network architecture, often via a ControlNet-style branched U-Net:

• ControlNet Branch: Per-object CNOCS maps—up to $x$6 per image—are input as auxiliary feature maps processed in parallel branches. These are merged into the main denoising UNet at multiple scales.
• Two-stage Training:

1. Supervised denoising: Minimize MSE between predicted velocity field and ground truth noise. 2. Reinforcement Learning Finetuning: On a data-balanced subset, optimize a pose-aligned reward comprising the 2D IoU error ($x$7) between predicted and ground-truth bounding boxes, and the KL divergence in the predicted orientation distribution ($x$8) (Qin et al., 20 Nov 2025).

This structure enforces strong alignment between synthesized images and pose control signals, leading to robust generalization and low error in both spatial localization and 3D orientation.

5. Comparison to NOCS and Related Representations

The CNOCS map extends the ideas of NOCS maps, which encode per-pixel canonical coordinates $x$9 in $x$0, usually for mesh-registered single-object scenarios (Kong et al., 15 Jun 2025). CNOCS relaxes the requirement for detailed meshes: any object that can be cuboid-approximated is eligible, eliminating the need for CAD