Class-Weighted-Aware Projection (CWAP)

• CWAP is a method that uses adjustable, class-dependent weights to prioritize critical semantic classes in projection operations such as 3D LiDAR range-view imaging.
• It introduces a simple weighting step during projection, adding negligible computation while resolving many-to-one mapping conflicts and improving sparse recovery.
• Empirical results, like a notable IoU boost for trucks in SemanticKITTI, demonstrate CWAP’s ability to enhance target class representation without significantly affecting overall metrics.

Class-Weighted-Aware Projection (CWAP) refers to a family of methods and algorithms that incorporate class prioritization into projection operations, either for resolving many-to-one conflicts in geometric data projections (such as LiDAR point cloud to range-view image), or for imposing class-weighted constraints in optimization (notably in sparse recovery and regularized learning). CWAP introduces adjustable, class-dependent weights so that decisions or projections can explicitly favor—or penalize—points, features, or variables according to their semantic or statistical class membership.

1. CWAP in Range-View Projection of 3D LiDAR Point Clouds

In the context of 3D LiDAR point cloud processing, range-view projection transforms the raw point set $\{\mathbf{p}_i = (x_i, y_i, z_i)\}_{i=1}^N$ into a 2D range image. A core challenge is the many-to-one mapping, where multiple points may fall into the same pixel; standard protocols resolve this by taking the point with minimum depth (smallest $d_i = \|\mathbf{p}_i\|_2$). Such depth-based rules neglect semantic class information and can suppress critical, yet non-central, object points (Mousavi et al., 26 Jan 2026).

CWAP generalizes this process by modifying the selection criterion: for each point $i$ of semantic class $\text{class}_i$, the projected score is

$s_i = d_i \cdot \frac{1}{w_i + \varepsilon}$

where $w_i$ is a user-chosen weight for $\text{class}_i$ and $\varepsilon > 0$ ensures numerical stability. The point with the lowest $s_i$ is selected per pixel.

Key properties:

• Large, positive $w_c$ decreases $s_i$ for class $c$, effectively making those points more likely to be selected.
• Negative $w_c$ ensures that class $c$ points always win over nonnegative-weighted classes within the same pixel.

This mechanism can prioritize "high-priority" semantic classes, such as pedestrians or cyclists, ensuring their representation is maximized in the projected range-view. The original depth-based rule is a special case with all $w_c = 0$.

2. Algorithmic Implementation in Range-View Projection

The CWAP procedure for LiDAR range projection is as follows (Mousavi et al., 26 Jan 2026):

1. For each point $i$, compute $d_i = \|\mathbf{p}_i\|_2$.
2. Project $\mathbf{p}_i$ to pixel $(u, v)$ via spherical/cylindrical mapping.
3. Retrieve $w_i = w[\text{class}_i]$ and compute $s_i = d_i / (w_i + \varepsilon)$.
4. Place $(i, s_i)$ in the candidate bucket for pixel $(u, v)$.
5. For each pixel, select $i^* = \arg\min s_i$ and assign $R(u, v) = d_{i^*}$ and associated features.
6. Output the completed range image.

This approach incurs negligible additional computation (one division and lookup per point) and requires minimal extra memory. Crucially, it is applied only during training since ground-truth class labels are not available at inference. At test time, conventional depth-based selection is used.

3. Weight Selection and Empirical Effects

CWAP’s flexibility arises from the weight vector $w = [w_1, ..., w_K]$, which encodes the relative importance of each semantic class. Heuristics for setting $w_c$ include:

• $w_c = 0$: leave class $c$ unweighted (default depth).
• $w_c > 0$: proportionally increase selection probability (e.g., $w_c=2$ doubles the effective priority).
• $w_c < 0$: class $c$ always wins in conflicts. In practice, non-target classes are set to zero, with positive or negative weights assigned to classes of interest.

Experimental evidence from SemanticKITTI using the RangeViT backbone demonstrates substantial improvements for targeted classes:

Projection	truck IoU	other-veh IoU	mIoU_inst	mIoU_stuff	overall mIoU
Depth	56.8	31.7	48.0	63.3	56.8
CAP	69.1	39.3	51.1	63.0	58.0
CWAP (-1)	77.5	33.9	51.4	63.0	58.1

Targeted classes (e.g., truck) exhibit a significant IoU increase (+20.7 pp for “truck”), with negligible impact on “stuff” and non-target instance classes (typically $<0.5$ pp). Over-weighting a single class can marginally degrade others due to the strict one-point-per-pixel constraint (Mousavi et al., 26 Jan 2026).

4. CWAP for Projection onto the Weighted $\ell_1$ Ball in Optimization

Class-Weighted-Aware Projection also manifests in the context of projected gradient methods under weighted $\ell_1$-norm constraints. Given $y \in \mathbb{R}^d$ and positive weights $w \in \mathbb{R}^d_+$, the goal is to solve: $$x^* = \arg\min_x \|x - y\|_2^2 \quad \text{s.t.} \quad \sum_{i=1}^d w_i|x_i| \leq a.$$ This admits a coordinate-wise soft-thresholding solution: $$x_i = \mathrm{sign}(y_i) \cdot \max\{ |y_i| - w_i \lambda^*,\ 0 \}$$ where $\lambda^*\geq 0$ is chosen to enforce the weighted $\ell_1$ constraint. In class-weighted regimes ($w_i = \alpha_{c(i)} w_i^{(base)}$ for class membership $c(i)$), this projection enforces group- or class-specific sparsity or penalty (Perez et al., 2020, Wang, 2015).

Three efficient algorithms have been proposed:

• w-sort: $O(d\log d)$, based on sorting.
• w-pivot$^f$: Near-linear in $d$ on average, quadratic in worst case, using partition and iterative refinement.
• w-bucket$^f$: Linear in both worst and average case, using radix-style bucketing and filtering.

Numerical benchmarks show that, for $d=10^7$, w-bucket$^f$ achieves $\sim$8 ms projection time on commodity CPU hardware (Perez et al., 2020).

5. Analytical Formulation and Solution in Weighted Simplex Problems

The weighted projection problem under a sum constraint and nonnegativity has an analogous structure: $$\min_{x \in \mathbb{R}^n} \frac{1}{2}\|x - y\|_2^2 + \sum_{i=1}^n w_i x_i \quad \text{s.t.} \quad \sum_{i=1}^n x_i = \tau,\; x_i \geq 0.$$ This is strictly convex and admits a unique minimizer. The KKT system yields: $$x_i = \max\left\{ y_i - w_i - \alpha,\ 0 \right\} \quad \text{with} \quad \sum_i x_i = \tau$$ where

$$\alpha = \frac{S(\rho) - \tau}{\rho},\quad \rho = \max\{ k : u_{(k)} > (S(k) - \tau)/k \}$$

for $u_i = y_i - w_i$, sorted in descending order and $S(k) = \sum_{i=1}^k u_{(i)}$ (Wang, 2015). When $w_i$ encodes a class cost, the projection yields class-weighted probability distributions or feature assignments.

6. Practical Considerations and Limitations

CWAP in range-view projection introduces minimal runtime and memory costs (a single division and lookup per point in LiDAR contexts). It fundamentally relies on the availability of class labels and thus cannot be directly used for inference where only observed data is available. Excessive weighting of a single class may saturate its representation in projected images, producing a minor negative "ripple" effect in the metrics of non-target classes due to exclusivity in pixel assignment. Overly large weights (above $2$–$3\times$) may induce diminishing returns or slight drops in overall accuracy (Mousavi et al., 26 Jan 2026).

In convex optimization settings, all CWAP projection algorithms discussed admit explicit space and time complexity guarantees, with w-bucket$^f$ preferred for large-scale tasks. Practical deployment demands careful attention to class-weight normalization, numerical stability, and in-place array operations for high-throughput applications (Perez et al., 2020).

7. Applications and Impact

CWAP’s utility is pronounced in:

• Semantic segmentation of 3D point clouds, especially where recovery of underrepresented classes (e.g., vulnerable road users) is safety-critical (Mousavi et al., 26 Jan 2026).
• Sparse feature selection and compressed sensing where structured or class-prioritized sparsity is desired (Perez et al., 2020).
• Adjustment of predicted class proportions or rankings with explicit reweighting (e.g., fair ranking, domain adaptation pipelines) (Wang, 2015). A plausible implication is that CWAP provides a generic tool for imposing soft or hard selection priorities wherever class semantics can be defined and leveraged in the projection or selection process.

---

References:

• (Mousavi et al., 26 Jan 2026): "Contextual Range-View Projection for 3D LiDAR Point Clouds"
• (Perez et al., 2020): "Efficient Projection Algorithms onto the Weighted l1 Ball"
• (Wang, 2015): "An $\mathcal{O}(n\log n)$ projection operator for weighted $\ell_1$-norm regularization with sum constraint"