L1RA: Rotation Averaging & Rank Assignment

• L1RA is an overloaded acronym representing L1 Rotation Averaging in photogrammetry and L1-regularized Rank Assignment in LoRA fine-tuning.
• In photogrammetry, L1 Rotation Averaging recovers absolute camera orientations from pairwise relative rotations to enable robust 3D reconstruction.
• In LLM fine-tuning, the method optimizes adapter ranks using an L1 penalty to prune and redistribute parameters under a fixed rank budget.

Searching arXiv for the cited L1RA papers and closely related references. L1RA is an overloaded acronym in the arXiv literature. In close-range photogrammetry, it denotes L1 Rotation Averaging, a method that takes as input a set of pairwise relative camera rotations and recovers a consistent set of absolute camera orientations $\{R_1,\dots,R_n\}$ before triangulation of a sparse point cloud (Utomo et al., 2017). In parameter-efficient fine-tuning of LLMs, it denotes L1-regularized Rank Assignment in LoRA Fine-Tuning, a technique that uses $\ell_1$ regularisation on adapter gating vectors to prune redundant ranks and redistribute them across adapters under a fixed rank budget (Singh et al., 5 Sep 2025). The two usages are algorithmically unrelated, but both place an $L_1$ penalty or cost at the center of the method. This suggests a shared naming logic based on robustness or sparsity rather than a shared application domain.

1. Polysemy and terminological scope

The term L1RA does not identify a single canonical method across fields. In the 2017 photogrammetry paper, the acronym is explicitly expanded as “L1 Rotation Averaging,” and its function is geometric pose recovery inside a reconstruction pipeline for temples in the Special Region of Yogyakarta (Utomo et al., 2017). In the 2025 LoRA paper, the same acronym is used for a method that dynamically distributes adapter rank during fine-tuning of quantized LLMs (Singh et al., 5 Sep 2025).

A nearby source of ambiguity is the use of L1 as shorthand for first-line treatment in clinical evidence synthesis. In the rheumatoid arthritis study on biologic therapies, “L1” refers to first-line biologic therapy effects in a bivariate network meta-analysis, not to L1RA as an algorithmic label (Bujkiewicz et al., 2022). For bibliographic and technical interpretation, the acronym therefore requires domain-specific disambiguation.

2. L1 Rotation Averaging as a camera-orientation estimator

In the photogrammetric usage, L1RA is introduced as the stage that follows feature matching and essential-matrix decomposition. Its input is the set of pairwise relative camera rotations, and its output is a consistent set of absolute orientations $\{R_1,\dots,R_n\}$. Once camera poses are available, 3D points can be triangulated to yield a sparse point cloud (Utomo et al., 2017).

The paper does not reproduce L1RA’s objective in detail and instead cites earlier literature. A common formulation of L1 rotation averaging is

$$\min_{R_1,\dots,R_n\,\in\,SO(3)} \sum_{(i,j)\in \mathcal{E}} \big\|\,R_{ij} \;-\; R_j\,R_i^\top \big\|_{1}$$

where $R_{ij}\in SO(3)$ is the measured relative rotation from camera $i$ to $j$, $R_i\in SO(3)$ are the unknown absolute rotations, $\|\cdot\|_1$ denotes the entry-wise $\ell_1$0 norm, and $\ell_1$1 is the set of image-pair edges for which a relative rotation has been estimated. A more robust variant replaces the matrix-entry $\ell_1$2 by the geodesic $\ell_1$3 on the Lie algebra:

$\ell_1$4

subject to each $\ell_1$5. The extracted summary explicitly notes that neither the paper nor its figures give these equations, and that they are included only to show the form of a typical L1RA problem (Utomo et al., 2017).

3. Position within the Yogyakarta temple reconstruction pipeline

The temple-reconstruction study collected 1319 photos from five temples in Yogyakarta. Using A-KAZE algorithm, keypoints of each image were obtained; LIOP was then employed to create feature descriptor from the detected points. After performing feature matching, L1RA was utilized to create sparse point clouds. To generate the geometry shape, MVS was used. Finally, FSSR and Large Scale Texturing were employed to deal with the surface and texture of the object (Utomo et al., 2017).

At the algorithmic level, the paper describes L1RA only at a high level: build a rotation-measurement graph from matched features and pairwise relative rotations $\ell_1$6, solve a global optimization using an $\ell_1$7 cost to mitigate the influence of outlier relative-rotation estimates, extract the final camera orientations $\ell_1$8, and triangulate matched feature points across suitably posed cameras to produce a sparse 3D point cloud. The authors state that L1RA “significantly outperforms other methods” because the $\ell_1$9 cost is more robust to spurious measurements. They do not report pseudocode, internal parameters, convergence tolerances, maximum iterations, or reweighting schedules; they state only that they used the implementation supplied in the OpenMVG library (Utomo et al., 2017).

The paper also does not provide an ablation or isolated timing or accuracy study for L1RA. All concrete performance figures concern the end-to-end reconstruction pipeline and the final model quality, measured by comparing the 3D images of the model with the original photos utilizing SSIM. On that end-to-end criterion, the method was reported to be on par with commercial methods such as PhotoModeler and PhotoScan (Utomo et al., 2017). No temple-specific limitations of L1RA are discussed; it is treated as a drop-in replacement for other rotation-averaging schemes.

4. L1RA as dynamic rank assignment in LoRA fine-tuning

In the LoRA literature, L1RA addresses a different problem: standard LoRA chooses a fixed rank $L_1$0 for every adapter, which can be sub-optimal if different layers or matrices require different capacity to adapt. The paper positions prior adaptive variants such as ADALORA, SORA, and DYLORA as methods that prune unused ranks but do not re-assign freed ranks to other adapters, so to reach a better allocation they must start from an inflated rank, which blows up memory (Singh et al., 5 Sep 2025).

L1RA instead starts with a total rank budget

$L_1$1

and adds a small gating vector $L_1$2 per adapter. For a pre-trained weight matrix $L_1$3 in layer $L_1$4, LoRA adds

$L_1$5

and L1RA computes

$L_1$6

Let $L_1$7 collect all trainable parameters $L_1$8, and let $L_1$9 be the downstream loss. The constrained optimization is

$\{R_1,\dots,R_n\}$0

An equivalent view is the Lagrangian

$\{R_1,\dots,R_n\}$1

together with the requirement that the active rank

$\{R_1,\dots,R_n\}$2

never lets $\{R_1,\dots,R_n\}$3. The paper states that $\{R_1,\dots,R_n\}$4 regularization is chosen because it drives individual gate components $\{R_1,\dots,R_n\}$5 exactly to zero, which is the cheapest way to detect and remove a rank dimension (Singh et al., 5 Sep 2025).

5. Optimization procedure, empirical characterization, and diagnostic signals

The high-level procedure begins by choosing an initial per-adapter rank $\{R_1,\dots,R_n\}$6 such that $\{R_1,\dots,R_n\}$7, initializing $\{R_1,\dots,R_n\}$8, $\{R_1,\dots,R_n\}$9, and $$\min_{R_1,\dots,R_n\,\in\,SO(3)} \sum_{(i,j)\in \mathcal{E}} \big\|\,R_{ij} \;-\; R_j\,R_i^\top \big\|_{1}$$0. Each batch computes the forward pass using $$\min_{R_1,\dots,R_n\,\in\,SO(3)} \sum_{(i,j)\in \mathcal{E}} \big\|\,R_{ij} \;-\; R_j\,R_i^\top \big\|_{1}$$1, optimizes $$\min_{R_1,\dots,R_n\,\in\,SO(3)} \sum_{(i,j)\in \mathcal{E}} \big\|\,R_{ij} \;-\; R_j\,R_i^\top \big\|_{1}$$2, prunes every coordinate with $$\min_{R_1,\dots,R_n\,\in\,SO(3)} \sum_{(i,j)\in \mathcal{E}} \big\|\,R_{ij} \;-\; R_j\,R_i^\top \big\|_{1}$$3 by removing the corresponding column of $$\min_{R_1,\dots,R_n\,\in\,SO(3)} \sum_{(i,j)\in \mathcal{E}} \big\|\,R_{ij} \;-\; R_j\,R_i^\top \big\|_{1}$$4 and row of $$\min_{R_1,\dots,R_n\,\in\,SO(3)} \sum_{(i,j)\in \mathcal{E}} \big\|\,R_{ij} \;-\; R_j\,R_i^\top \big\|_{1}$$5, and then redistributes freed ranks by selecting the adapter whose minimum gate value $$\min_{R_1,\dots,R_n\,\in\,SO(3)} \sum_{(i,j)\in \mathcal{E}} \big\|\,R_{ij} \;-\; R_j\,R_i^\top \big\|_{1}$$6 is largest. A new column, row, and gate value are appended to that adapter, and optional normalization rescales each $$\min_{R_1,\dots,R_n\,\in\,SO(3)} \sum_{(i,j)\in \mathcal{E}} \big\|\,R_{ij} \;-\; R_j\,R_i^\top \big\|_{1}$$7 so that $$\min_{R_1,\dots,R_n\,\in\,SO(3)} \sum_{(i,j)\in \mathcal{E}} \big\|\,R_{ij} \;-\; R_j\,R_i^\top \big\|_{1}$$8 while preserving the sparsity pattern (Singh et al., 5 Sep 2025).

The reported fine-tuning scenario is instruction-tuning a chat-assistant on the OPENORCA dataset, using MISTRAL 7B v0.3 and LLAMA 3.1 8B, both 4-bit quantized. Baselines are vanilla LoRA with rank $$\min_{R_1,\dots,R_n\,\in\,SO(3)} \sum_{(i,j)\in \mathcal{E}} \big\|\,R_{ij} \;-\; R_j\,R_i^\top \big\|_{1}$$9, ADALORA starting from $R_{ij}\in SO(3)$0 targeting $R_{ij}\in SO(3)$1, and ADALORA starting at $R_{ij}\in SO(3)$2 targeting $R_{ij}\in SO(3)$3. Metrics are test perplexity (PPL), total fine-tuning time, GPU memory peak, and # trainable adapter parameters (Singh et al., 5 Sep 2025).

System	Configuration	Reported metrics
LLAMA 3.1 8B	LoRA $R_{ij}\in SO(3)$4	PPL=3.32; time≈31 k s; mem=13.84 GB; params=41.9 M
LLAMA 3.1 8B	ADALORA $R_{ij}\in SO(3)$5	PPL≈3.63 (+9%); time≈32.98 k s (+6%); mem≈14.00 GB (+1%); params≈62.9 M (+50%)
LLAMA 3.1 8B	ADALORA $R_{ij}\in SO(3)$6	PPL≈3.57 (+7.5%); time≈32.96 k s (+6.3%); mem≈13.76 GB (−0.6%); params≈42.0 M (+0.1%)
LLAMA 3.1 8B	L1RA $R_{ij}\in SO(3)$7	PPL=3.25 (−2.1%); time=31.25 k s (+0.8%); mem=14.23 GB (+2.8%); params=45.2 M (+7.7%)
MISTRAL 7B v0.3	LoRA	PPL=2.93; time=37.9 k s; mem=13.58 GB; params=41.9 M
MISTRAL 7B v0.3	ADALORA $R_{ij}\in SO(3)$8	PPL=3.16 (+7.9%); time=40.23 k s (+6.2%); mem=13.82 GB (+1.8%); params=62.9 M (+50%)
MISTRAL 7B v0.3	ADALORA $R_{ij}\in SO(3)$9	PPL=3.16 (+7.9%); time=40.22 k s (+6.1%); mem=13.59 GB (+0.1%); params=42.0 M (+0.1%)
MISTRAL 7B v0.3	L1RA $i$0	PPL=2.91 (−0.7%); time=37.97 k s (+0.2%); mem=13.94 GB (+2.6%); params=50.1 M (+19.4%)

The paper’s stated takeaway is that L1RA matches or improves over vanilla LoRA in perplexity while keeping within the original memory footprint, avoids the large parameter-count blow-up of ADALORA, and incurs only a few percent of extra VRAM. Its post-training rank-distribution analysis reports that the highest ranks are consistently assigned to the three sub-matrices of the Transformer feed-forward network—$i$1, $i$2, and $i$3—while the attention matrices $i$4, $i$5, $i$6, and $i$7 receive lower ranks, and ranks tend to increase in layers closer to the final output. The paper states that this suggests that adapting high-level FFN transforms is more critical for instruction-tuning than reshaping self-attention at every layer (Singh et al., 5 Sep 2025).

6. Comparative interpretation, limitations, and recurrent misconceptions

The two meanings of L1RA should not be conflated. In photogrammetry, the method estimates elements of $i$8 from a rotation-measurement graph and enables triangulation of sparse 3D structure before later stages such as MVS, FSSR, and Large Scale Texturing (Utomo et al., 2017). In LoRA fine-tuning, the method operates on trainable low-rank adapters $i$9 and reallocates active rank under a global budget $j$0 (Singh et al., 5 Sep 2025). The commonality is methodological emphasis on an $j$1 criterion: outlier resistance in one case, exact sparsification of gate coordinates in the other.

A common misconception is to read L1RA as a stable cross-domain acronym. The rheumatoid arthritis paper illustrates why this is unreliable: there, L1 denotes first-line therapy, with covariate-adjusted logistic regression estimates $j$2 for first-line contrasts entering a bivariate network meta-analysis, but no L1RA algorithm is involved (Bujkiewicz et al., 2022). Another misconception is to overgeneralize reported performance. The temple paper provides no ablation or isolated timing or accuracy study of L1RA itself, and its concrete evaluation concerns the end-to-end reconstruction pipeline (Utomo et al., 2017). The LoRA paper reports validation on two 8 billion-parameter models and a single instruction-tuning task, and explicitly identifies further benchmarks across scale, tasks, and architectures as a valuable next step (Singh et al., 5 Sep 2025).

Taken together, the available literature supports a narrow and technically specific reading of each usage. L1 Rotation Averaging is a globally optimal, outlier-robust pose-estimation component in close-range photogrammetry, whereas L1-regularized Rank Assignment is a budget-constrained adapter-allocation strategy for efficient LLM fine-tuning. Any interpretation of “L1RA” therefore depends primarily on the surrounding research domain, the objects being optimized, and the role of the $j$3 term in the objective.