acceptodds
Under review as a conference paper at ICLR 2027

SOMo: Unified Spatial and Observation-Grounded Scene Memory

Abstract

Questions about a robot's past observations often identify an object through its relationships to other objects and the context of a recorded encounter. An agent answering such questions through separate searches must assemble compatible objects and observation evidence from intermediate results; an early selection can exclude the intended answer and misdirect subsequent tool calls. We introduce SOMo, a scene memory that lets spatial relationships and observation context guide object selection during retrieval. Through a shared interface, the agent specifies object descriptions, required relationships, and spatial and temporal restrictions. SOMo jointly selects compatible objects and returns their mapped locations with the requested observation evidence, supporting proximity, co-observation, and robot-relative bearing queries. A sparse spatial hierarchy connects object groups to recorded robot visits and summarizes their spatial extent and observation intervals, allowing constraints to eliminate incompatible groups before individual verification. Adaptive planning selects the grouping resolution and constraint order using execution-cost estimates and counts of the object pairs remaining after grouped filtering, before expanding those pairs. Joint semantic ranking then selects among the verified combinations. With Qwen3.5-9B for both methods, SOMo reduces positional and temporal error relative to DAAAM by 45.3% and 37.6% on NaVQA, and by 68.09% and 64.32% on CODaAG's category averages. On SG3D, SOMo with Qwen3.5-9B achieves 31.82% sub-task grounding accuracy, compared with 22.16% for DAAAM with GPT-5-mini. On CODaAG, average tool use falls from 5.62 to 1.46 calls and answering time from 31.26 to 6.29 seconds. Separate candidate-generation benchmarks show 7–25 speedups over blocked NumPy scans on CODa and 16–193 on KITTI-360 with identical candidate results.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Scene Retrieval, LLM Agent

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_88qwOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper presents SOMo, a scene memory for retrieving objects from a robot’s past observations using semantic, spatial and temporal information. The agent makes a unified "search_objects" call with object descriptions and a relation, and the backend does the filtering with a hierarchical spatial memory, observation associations and an adaptive planner. Memory construction itself is from DAAAM. Experiments on several QA and grounding benchmarks show lower error and fewer tool calls than previous methods.

Critical strengths

The problem of retrieving contextual information from long-term robot memory is relevant, and handling relational constraints inside retrieval rather than across LLM tool calls is a reasonable idea. The reduction in tool calls seems plausible.

Critical weaknesses
  1. Unclear problem formulation and contribution. The paper does not clearly formulate the problem it addresses, and the contribution is not clearly specified at the end of the introduction. There are many components stacked on top of DAAAM, but it is not clear which of them is essential or new, or why the whole system needs to be this complicated. A lot of it looks like standard spatial indexing, and several design choices are stated without much motivation. SOMo also changes the retrieval interface and backend at the same time, without an ablation separating their effects, so it is difficult to tell what is actually responsible for the reported gains. Overall, I found the method and the paper hard to follow.

  2. Quality control of the manuscript. The eMEM reference lists affiliations as author names (and invents a first name for the first author), and several other references contain conflicting publication years. The authors disclose AI assistance for implementation and writing, which is fine by itself, but it does make me wonder whether the generated content was checked carefully.

Questions for authors
  1. Could the authors state the problem formulation precisely and say what they consider the main contribution?
  2. What does DAAAM plus a simple joint-search tool achieve, without the hierarchy and planner? More generally, which components are actually responsible for the gains?
  3. How does SOMo perform on the generated benchmarks when the agent is not explicitly told which retrieval relation or tool to use?
2 comments

W1 / Q1. We thank the reviewer for pointing out that the problem formulation and contribution need to be stated more clearly. Given a memory of mapped objects, semantic features, observation histories, and robot trajectory samples, the problem is to retrieve an ordered assignment of distinct objects to the descriptions in a request, subject to spatial and observation constraints. For example, “the bike near a bench that you saw together between time A and B” requires a bike–bench assignment satisfying both proximity and common observation support within the requested period. Among feasible assignments, SOMo ranks the combinations by their semantic match to all requested roles and returns the selected objects together with supporting observation evidence.

The main contribution is contextual retrieval that jointly selects objects and their relevant encounters under a shared set of conditions. Separate object searches leave the agent to assemble and verify combinations after retrieval, and an early choice of a reference object can exclude the correct target. SOMo moves these checks into retrieval, allowing the relationship and observation context to influence which objects are selected. Its joint ranking prioritizes the weakest semantic role match, while returned observation evidence supports subsequent answer generation. The sparse hierarchy, object–visit associations, and adaptive planner provide an efficient implementation of these retrieval operations. We will revise the introduction and method overview to distinguish the retrieval formulation and its effect on answer quality from the execution mechanisms and their effect on efficiency.

W2. We thank the reviewer for identifying this issue. We collected all BibTeX entries manually, and the incorrect eMEM entry may have originated from an upstream metadata error in Semantic Scholar at the time of collection. We will correct the affected references, including the author information and publication-year inconsistencies, in the manuscript.

Q2. DAAAM with a simple pair-search tool improves over DAAAM, but remains below SOMo. The following NaVQA results use Qwen3.5-9B throughout.

Table 1. DAAAM with an explicit object-pair search tool compared with SOMo.

MethodQuestion accuracy ↑Position error (m) ↓Temporal error (min) ↓
DAAAM0.67034.3601.613
DAAAM + pair-search tool0.67227.4691.524
SOMo0.71818.7921.007

DAAAM equipped with our full unified search tool—including joint constraint verification, role assignment and ranking, and returned observation evidence—effectively becomes SOMo at the retrieval level, since SOMo already uses DAAAM's memory-construction pipeline. The hierarchy and planner accelerate execution of this retrieval contract. Replacing them with exhaustive evaluation over the same records and constraints, while preserving ranking and returned evidence, preserves the retrieval results and therefore the information available to the answering agent. We do not attribute task-accuracy gains to the hierarchy or planner.

To isolate the value of retrieving multiple objects, we also restrict SOMo to single-object queries while holding everything else fixed. Performance decreases substantially.

Table 2. NaVQA ablation restricting SOMo to single-object queries, with Qwen3.5-9B throughout.

MethodQuestion accuracy ↑Position error (m) ↓Temporal error (min) ↓
DAAAM0.67034.3601.613
SOMo, single-object queries only (k = 1)0.70226.7811.135
SOMo0.71818.7921.007

We further separate ranking and evidence materialization using 105 joint-object CODaAG questions. All arms share an exhaustive backend, fixed requests, and identical prompts; answer generation proceeds without agent-driven query selection or follow-up searches.

Table 3. Fixed-query ablation of joint ranking and returned evidence.

RankingReturned evidencePosition error (m) ↓Temporal error (min) ↓
MeanRaw support and trajectory13.760.189
LeximinRaw support and trajectory16.410.193
MeanRaw + materialized witnesses13.660.192
LeximinRaw + materialized witnesses10.830.134

Leximin combined with materialized observation witnesses gives the lowest errors in this study. Leximin alone does not improve over mean ranking with raw evidence, so these results support the combination of joint ranking and usable observation evidence rather than an independent benefit from every component. Together, these ablations distinguish the quality benefits of joint retrieval and evidence presentation from the computational benefits of indexing and planning.

Q3. The tools’ behavior and arguments are described in their docstrings and tool descriptions, which are presented to the agent. The agent constructs tool calls using this documentation and the question. Table 4 distinguishes this setting from using tool documentation alone. We found that removing explicit tool guidance slightly lowers SOMo's performance. However, the best call argument to use often remains quite identifiable from the question so the impact remains limited.

Table 4. CODaAG results across four query categories for SOMo with Qwen3.5-9B under different agent-guidance settings. Pos. is positional error (m); Temp. is temporal error (min); lower is better.

Agent guidanceSingle-Object Pos. ↓Single-Object Temp. ↓Multi-Object Co-Observation Pos. ↓Multi-Object Co-Observation Temp. ↓Multi-Object in Proximity Pos. ↓Multi-Object in Proximity Temp. ↓Robot-Relative Angle Pos. ↓Robot-Relative Angle Temp. ↓
Tool descriptions + tool guidance17.690.95222.821.16224.421.12425.261.213
Tool docstrings/descriptions only17.710.95223.251.16824.961.14125.311.219
Reviewer_bCbeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper presents SOMo, a scene memory framework for LLM agents that addresses the failure mode of separate-object retrieval by consolidating relational, spatial, and temporal constraints into a single retrieval interface. Operating over an object-centric mapping backbone, SOMo integrates a sparse multi-resolution Morton index, an adaptive query planner, member-level verification, and leximin-based role ranking to return multi-object role assignments with observation evidence.

Critical strengths

Effective Problem Formulation and Relational Retrieval Design: The core insight that multi-object LLM agents frequently fail when forced to commit early to reference objects through decoupled searches is well-founded. By unifying multi-role, spatial, and temporal constraints directly within a single retrieval primitive, SOMo effectively reduces early-commitment errors. This API-level shift provides a practical mechanism for relational reasoning in embodied AI agents, demonstrated by marked reductions in both execution latency and total tool calls.

Critical weaknesses

Decoupled Contributions and Backend Justification:

  • There is a fundamental misalignment between the paper's task-level quality gains and its technical indexing and planning contributions. Because the backend guarantees exact candidate verification, all task accuracy improvements stem entirely from the consolidated search_objects interface and ranking logic, not from the Morton hierarchy or adaptive planner. At the evaluated dataset scales (1.5k to 3.5k objects), brute-force NumPy evaluation takes roughly 5–50 ms on CODa and at most ~1.5 s on KITTI-360, small relative to 6–30 s end-to-end LLM response times, making backend speedups largely negligible in practice.
  • Furthermore, the adaptive planner is slower than a fixed observation-first policy on real-world combined queries (0.80×), and its optional challenger provides little aggregate benefit, performing at 0.80x relative to a simple observation-first heuristic. Without demonstrating necessity through large-scale scenarios or established spatial database baselines (e.g., PostGIS, R-trees, cKDTree), the extensive technical machinery in Section 3 lacks empirical justification.
Questions for authors
  • Can the authors confirm whether task accuracy remains identical when SOMo's backend is replaced with brute-force or KD-tree search, and evaluate DAAAM when equipped with an explicit pair-search tool to isolate the true benefit of the unified interface?

  • How sensitive are results to the distance thresholds the agent chooses, e.g., 10 m for “near,” or interpreting “in front of” as proximity in Appendix G?

  • How does the framework handle directional or support relations (such as "in front of" or "between") that are currently approximated by distance thresholds, and how much of the remaining SG3D gap is attributable to this limitation?

  • What happens when SOMo is restricted to single-object queries (k = 1), which would isolate the value of joint retrieval?

2 comments

W1. As our results demonstrate and the reviewer notes, retrieval latency can increase substantially with the number of objects and scene size. However, end-to-end evaluation at larger scales is currently constrained by the availability of datasets with sufficiently detailed object annotations to distinguish individual objects and groups. As robotic environments and datasets grow, we expect efficient retrieval to become increasingly important, and our approach provides a foundation for this setting.

W2. Our observation-first retrieval mechanism and adaptive planner are complementary components of SOMo. The former exploits object–visit associations to generate candidates, while the latter selects the execution order and spatial resolution according to the query. We acknowledge that planning overhead can outweigh its benefit when a fixed policy already suits the workload: observation-first is faster overall on our real-scene combined queries. Adaptive planning nevertheless improves aggregate performance over both fixed-order policies on the controlled workload and avoids expensive execution choices when query selectivity changes. These results support its use for varying workloads, while recognizing the efficiency of observation-first execution when it consistently matches the queries.

We additionally compare SOMo with standard spatial indexing and database baselines. SOMo integrates spatial constraints with trajectory and object-observation evidence through a unified retrieval interface. Because existing spatial databases cannot tackle such queries, we isolate spatial retrieval by asking each method to return all object pairs within a specified distance, using identical object positions and distance predicates. SOMo’s implementation significantly outperforms the other solutions.

Table 1. Median latency for spatial pair retrieval on the CODa scene and the KITTI-360 scene, using distance thresholds of 10 m and 100 m, respectively.

BackendCODa, 10 m (ms) ↓KITTI-360, 100 m (ms) ↓
NumPy brute force17.595176.037
cKDTree13.089321.563
R-tree28.128247.659
PostGIS34.174400.620
SOMo (Python)4.56052.507

Table 2. Number of objects and possible unique pairs for the reported measurements.

SceneObjectsPossible unique pairsPairs returned within radius
CODa seq. 00, 10 m1,5011,125,75033,892
KITTI-360 drive 0009, 100 m3,5566,320,790443,545

Q1. Our backend is meant to accelerate the precise answer given the specific request; therefore, the backend choice does not affect the returned answer, and the accuracy remains identical. We further evaluate DAAAM with an additional pair search tool without the rich unified interface that SOMo developed. We demonstrate the comparison below. As shown in the table, the joint pair search tool brings a significant reduction in position error. However, as the questions heavily involve different temporal aspects, DAAAM with a pair search tool cannot reach the performance of SOMo.

Table 3. NaVQA comparison with an added pair-search tool.

MethodQuestion accuracy ↑Position error (m) ↓Temporal error (min) ↓
DAAAM (Qwen3.5-9B)0.67034.3601.613
DAAAM (Qwen3.5-9B) with pair tool0.67227.4691.524
SOMo (Qwen3.5-9B)0.71818.7921.007

Q2. We explicitly intercept the agent's interpreted distance and set it to half or double its value in our backend. We found that increasing the distance usually does not significantly change the accuracy, while reducing it by half slightly lowers the performance. We address the "in front of" question in the next point.

MethodQuestion accuracy ↑Position error (m) ↓Temporal error (min) ↓
SOMo - half distance (Qwen3.5-9B)0.70821.1311.101
SOMo (Qwen3.5-9B)0.71818.7921.007
SOMo - double distance (Qwen3.5-9B)0.71619.1291.003

Q3. For scene questions that do not specify a viewing direction or reference frame, “in front of” can be ambiguous: an object in front of another from one perspective may appear behind it from another. In such cases, approximating the relation using distance proximity does not actually lose material information. “Between” conveys a more specific spatial arrangement, but this may provide limited additional benefit for object retrieval when other contextual clues already identify the target. There are limited examples of "between" in the SG3D dataset; specifically, one mentions "podium close to the dining table and between several chairs" and another mentions "chair between the desk and the speaker". In either case, SOMo was able to identify the correct target by using object tuples, so there is no particular performance loss due to not modeling "between" explicitly.

Q4. We additionally show SOMo's performance when the maximum number of object instances is kept at 1 while holding everything else fixed. We see substantially lower performance, indicating the key benefit of our unified design covering multiple objects.

Table 4. NaVQA ablation restricting SOMo to single-object queries (k = 1).

MethodQuestion accuracy ↑Position error (m) ↓Temporal error (min) ↓
DAAAM (Qwen3.5-9B)0.67034.3601.613
SOMo (Qwen3.5-9B), single-object queries (k = 1)0.70226.7811.135
SOMo (Qwen3.5-9B)0.71818.7921.007
Reviewer_E1xbOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces SOMo, a scene-memory system for retrieving objects jointly under spatial and observation constraints. Built on DAAAM’s object-centric mapping pipeline, SOMo combines a unified retrieval interface, a sparse multiresolution spatial hierarchy, object-visit associations, grouped constraint filtering, and adaptive query planning. Feasible object combinations are ranked through joint semantic assignment and returned with observation evidence. Experiments evaluate spatiotemporal question answering, contextual retrieval, and sequential grounding, alongside backend efficiency and execution-planning studies.

Critical strengths
  1. The proposed system demonstrates practical effectiveness. With matched reasoning models, SOMo reduces positional and temporal errors on NaVQA and improves sequential grounding accuracy on SG3D over DAAAM. These results provide encouraging evidence of the integrated system’s utility for contextual retrieval and downstream reasoning.
Critical weaknesses
  1. Unclear technical novelty over recent context-aware memory systems. The intro and Fig. 1 motivate SOMo mainly by contrasting it with separate per-object searches (like vanilla DAAAM), where picking a candidate too early can miss the right relational context. While that limitation is real for naive tool-calling workflows, it is much less clear what remains unsolved in the more recent context-aware systems cited in Sec. 2.1 (e.g., R4, SpatialMem, STaR, eMEM). The paper briefly mentions these works but does not compare in sufficient detail how they handle candidate generation, relational constraints, or object-role assignment, nor where their mechanisms fall short. SOMo puts together object-visit links, Morton-code hierarchical filtering, adaptive planning, and leximin ranking into a working pipeline, but it is hard to tell which part represents a genuine algorithmic advance over existing spatio-temporal memory designs.
  2. Missing recent baselines and incomplete ablations on accuracy and indexing. First, several closely related methods that were available well before the deadline are missing from the experiments. For instance, STaR on NaVQA (which is already cited in the paper and open-sourced in July 2026), Worth Remembering [1] and UQ-DAAAM [2] on OC-NaVQA, and FOUND-IT [3] on SG3D. Second, the ablations in Appendix C.4.4 only test execution planning (resolution and constraint order), leaving their individual contributions to the reported accuracy gains (Sec. 3.5) unquantified. Because the system also changes the tool interface and prompt guidance (Appendix C.2), we need a controlled ablation --- fixing the object records, observation support, prompts, and output format --- to see what leximin ranking and evidence materialization actually add over a simple joint-retrieval baseline. Finally, benchmarking the candidate kernels only against blocked NumPy brute-force scans (Tables 7-8) is a weak baseline; comparing against standard spatial-temporal index structures under the same predicates is necessary to justify the proposed hierarchy.

[1] Nicolas Gorlo, Derek K. Wise, Alberto Speranzon, and Luca Carlone. “Worth Remembering: Surprise-Gated Robot Episodic Memory.” arXiv:2606.03787.

[2] Harry Zhang, Nicolas Gorlo, and Luca Carlone. “Remember with Confidence: Uncertainty Quantification for Spatio-temporal Memory with Probabilistic Guarantees.” arXiv:2606.08277.

[3] Dominic Maggio, Nicolas Gorlo, Kris Hauser, and Luca Carlone. “FOUND-IT: Foundation-model-first Task-driven 3D Scene Graphs with Granularity on Demand.” arXiv:2605.25371.

Questions for authors

Please refer to W1-W2 to clarify the technical novelty and more comparable baselines of the submission.

2 comments

W1

We illustrate the behavioral differences between SOMo, STaR, and eMEM through one example. Without open-source implementations of R4 and SpatialMem, we cannot compare their execution traces.

We compare the methods on the question: "Where is the statue in front of the Bill and Melinda Gates Computer Science building?"

The STaR trace comes from our best local reproduction attempt, whose aggregate performance remains far below the reported results. It illustrates retrieval behavior in this run, rather than establishing comparative benchmark performance.

STaR makes three text-retrieval calls:

  1. Searches the complete relational phrase. It calls retrieve_from_text with the full statue-and-building description. The leading caption, at 10:56:26, describes a person before a glass door labeled “Bill & Melinda Gates Computer Science.” Another describes the building's exterior. Neither identifies the statue.

  2. Receives a potentially useful sculpture observation. The same retrieval returns a caption at 10:55:53 describing “a large white structure, which appears to be a sculpture or monument.” The evidence thus includes both the named building and a possible sculpture, but as separate caption records with observation times and robot positions. It does not assign target and reference roles or establish their spatial relationship.

  3. Searches the building name separately. It calls retrieve_from_text with Bill and Melinda Gates Computer Science building. Results again describe the person at the labeled door and the building's surroundings, without resolving which sculpture belongs with it.

  4. Repeats the complete relational search. The third call repeats the original phrase and returns the same leading building-related captions. The agent still does not identify the target sculpture or verify its association with the building.

  5. Returns its current robot position. The final answer, [-88.93, -203.63, 9.81], is exactly the supplied current robot location. The agent uses its presence near the building to answer without selecting a statue or linking this position to the sculpture observation.

The retrieval contains both the building name and a sculpture description. The unresolved step is binding the target sculpture to the reference building and choosing the corresponding encounter. STaR makes relevant evidence available, but repeating the semantic search does not establish that binding in this execution.

eMEM: A building-related memory supplies the statue's answer

eMEM takes a shorter path:

  1. Treats the complete description as one concept. The agent calls:

    {
      "tool": "locate",
      "concept": "statue in front of the Bill and Melinda Gates Computer Science building",
      "n_results": 1
    }
    
  2. Selects one matching memory. The caption describes a person before the glass door carrying the building's name. It identifies the reference-building context but no statue.

  3. Converts that memory into a location. locate computes a centroid and spread from matching memories. With n_results=1, the centroid is the selected memory's position: (-77.6, 205.6, -2.0), with a 1 m radius and “Based on: 1 memories.” The radius is the single-memory default, not a verified statue–building distance or localization confidence.

  4. Accepts the location without further verification. The agent returns this location without another query to identify a statue or verify its relationship to the building. Its observer-position error is 22.93 m.

Here, a memory matching the reference building becomes the answer to a target-statue question. The semantic lookup has no separate requirement that objects fill the statue and building roles, so a building-related caption can match while leaving the target unidentified. eMEM retains entity associations and other retrieval tools, but this locate operation does not use them to verify a joint assignment.

In these executions, STaR retrieves relevant evidence and eMEM resolves a semantic concept to a location, but neither establishes the target–reference association before answering. As demonstrated in Appendix G.1, SOMo incorporates that association into retrieval and explicitly returns the resulting candidates and evidence.

W2 – Part 1

We study the detailed ablation raised by the reviewer in W2 through a fixed-query ablation on 105 joint-object CODaAG questions. All arms share an exhaustive search backend, fixed requests, and identical prompts. We vary ranking and returned evidence, then evaluate answer generation without agent-driven query selection or follow-up searches. As demonstrated, the top choice from Leximin, combined with materialized witnesses, shows the highest performance.

Table 1. Fixed-query ablation of joint ranking and evidence materialization.

RankingReturned evidencePosition error (m)Time error (min)
MeanRaw support and trajectory13.760.189
LeximinRaw support and trajectory16.410.193
MeanRaw + materialized witnesses13.660.192
LeximinRaw + materialized witnesses10.830.134

W2 – Part 2

We thank the reviewer for identifying these relevant comparisons. Below, we place the authors’ reported results alongside SOMo’s paper results.

Table 2. Reported question-answering results. NaVQA entries show Short / Medium / Long contexts; OC-NaVQA entries show All. SOMo means are shown without error bars.

BenchmarkMethod and reasoning modelQuestion/text accuracy ↑Position error (m) ↓Temporal error (min) ↓
NaVQASTaR, GPT-4.1-mini0.82 / 0.84 / 0.804.2 / 10.8 / 16.90.4 / 0.9 / 2.7
NaVQASOMo, GPT-5-mini0.724 / 0.705 / 0.7906.704 / 16.180 / 25.8490.365 / 0.683 / 1.501
NaVQASOMo, Qwen3.5-9B0.714 / 0.689 / 0.7805.013 / 17.750 / 27.6240.382 / 0.654 / 1.554
OC-NaVQAWorth Remembering, GPT-5-mini0.79636.571.510
OC-NaVQAUQ-DAAAM, GPT-5-mini0.76137.841.589
OC-NaVQASOMo, GPT-5-mini0.75230.891.318
OC-NaVQASOMo, Qwen3.5-9B0.73431.531.362

On OC-NaVQA, Worth Remembering and UQ-DAAAM report higher question accuracy, while SOMo reports lower positional and temporal errors. Their improvements to episode selection and object descriptions address complementary aspects of memory construction.

Table 2 reports STaR’s published results, since our local reproduction did not recover its reported performance.

Regarding FOUND-IT, it reports 39.7% subtask accuracy and 19.0% task accuracy on SG3D, stating that it follows the ASHiTA/DAAAM protocol, which we also follow. However, its evaluation description specifies 1,000 queries per scene and distance-threshold scoring, whereas ASHiTA describes 205 tasks across eight scenes and centroid-within-bounding-box scoring. This discrepancy leaves the exact protocol alignment unclear

W2 – Part 3

To address the concern about benchmarking only against blocked NumPy scans, we additionally compare SOMo with cKDTree, R-tree, and PostGIS. We isolate spatial pair retrieval: each method returns all object pairs within a specified distance, using identical object positions and distance predicates. This evaluates the spatial component of SOMo's retrieval backend under matched queries.

Table 3. Median latency for spatial pair retrieval on CODa and KITTI-360, using distance thresholds of 10 m and 100 m, respectively.

BackendCODa, 10 m (ms) ↓KITTI-360, 100 m (ms) ↓
NumPy brute force17.595176.037
cKDTree13.089321.563
R-tree28.128247.659
PostGIS34.174400.620
SOMo (Python)4.56052.507

Table 4. Scene sizes and pair populations for the reported measurements.

SceneObjectsPossible unique pairsPairs returned within radius
CODa seq. 00, 10 m1,5011,125,75033,892
KITTI-360 drive 0009, 100 m3,5566,320,790443,545

SOMo achieves lower median latency than all compared implementations on both tested scenes. These results supplement the NumPy comparison with established spatial indexing and database baselines, supporting the efficiency of SOMo's spatial retrieval implementation under the evaluated predicates.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.