{"id":"60d039b4-4c2a-4932-aee6-d622ca1fa780","arxiv_id":"2606.23113","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":2.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"Ensemble of three foundation encoders with Mask2Former and per-class validation weighting achieves second place at 75.40% mIoU on GOOSE 2D fine-grained segmentation, attributing gains mainly to pretraining diversity.","lead":"This paper reports an ensemble of DINOv3, SigLIP2, and InternImage encoders paired with a Mask2Former decoder that reaches 75.40% mIoU on the GOOSE outdoor segmentation challenge. A smart generalist might read it to see how pretraining differences and simple weighting can improve real-world scene parsing without needing larger models.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Central claim that pretraining recipe dominates parameter count and decoder design lacks any controlled ablations or comparisons.","rationale":"The reader's note that the pretraining-dominance claim lacks ablations directly matches the load-bearing gap for the strongest claim. The ensemble-complementarity concern is a secondary risk that only becomes relevant once the primary claim is evidenced.","tokens_in":1672,"tokens_out":328,"duration_ms":17913,"concrete_test":"Add a table reporting per-encoder validation mIoU for each of DINOv3, SigLIP2 and InternImage using the identical Mask2Former decoder, training schedule and crop size; also include at least one run with a different decoder (e.g., DeepLabv3+) on one encoder. If the spread attributable to pretraining is not substantially larger than the decoder or size effects, the dominance claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim requires evidence that pretraining objective produces larger accuracy deltas than model size or decoder choice when other factors are held fixed. The report describes only a single ensemble (DINOv3 + SigLIP2 + InternImage, all paired with Mask2Former, trained with the same long schedule/EMA/crop/TTA recipe) and its test mIoU; no tables compare (a) the three encoders individually under identical decoder/training, (b) parameter-matched variants with different pretraining, or (c) the same encoder with an alternative decoder. Without these isolations the dominance statement remains an assertion rather than a result.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"This technical report describes a solution for the ICRA 2026 GOOSE 2D Fine-Grained Semantic Segmentation Challenge that ensembles three foundation encoders (DINOv3, SigLIP2, InternImage) paired with Mask2Former decoders. The encoders are selected for complementary pretraining objectives and combined via per-class validation-IoU weighting; the system is trained with long schedules, EMA, large crops, and multi-scale/flip TTA. The submission reports 75.40% composite mIoU on the official held-out test set and second place in the challenge. The authors further claim that encoder pretraining recipe, rather than parameter count or decoder design, is the dominant accuracy factor on this benchmark.","tokens_in":1800,"tokens_out":577,"duration_ms":15251,"significance":"The reported test-set result is a concrete, externally validated performance point on a challenging 56-class outdoor segmentation task across multiple camera platforms. If the pretraining-dominance interpretation were supported by controlled comparisons, the work would usefully highlight the value of pretraining diversity in ensembles for robust scene understanding. As written, the empirical result stands but the interpretive claim does not add substantial new insight beyond the leaderboard placement.","major_comments":[{"comment":"Abstract: the central claim that 'the encoder's pretraining recipe, rather than its parameter count or the decoder design, is the dominant factor' is unsupported. The manuscript presents only a single ensemble configuration (all three encoders with Mask2Former under identical training/TTA) and its test mIoU; no tables or sections isolate pretraining effects while holding model size and decoder fixed, nor compare the three encoders individually under matched conditions.","section":"Abstract"},{"comment":"The complementarity assumption underlying the ensemble (distinct pretraining objectives plus per-class IoU weighting) is stated but not tested for robustness. No ablation shows that removing any one encoder or altering the weighting scheme materially changes test performance, leaving open whether the reported gain is due to pretraining diversity or simply to ensembling three strong models.","section":"Abstract / method description"}],"minor_comments":[{"comment":"Abstract and results section: no error bars, run-to-run variance, or statistical tests accompany the 75.40% mIoU figure, making it difficult to assess whether the second-place margin is reliable.","section":"Abstract"},{"comment":"The manuscript would benefit from a short table listing the three encoders' individual parameter counts, pretraining datasets/objectives, and (if available) their standalone validation mIoU under the shared training recipe.","section":"method / experiments"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments on the strength of our interpretive claims. We agree that the manuscript does not contain the controlled experiments needed to support the stated conclusions about pretraining dominance and ensemble complementarity, and we will revise the text accordingly.","responses":[{"response":"We agree that the central claim is unsupported by the experiments in the manuscript, which reports only the final ensemble result. We will revise the abstract to remove this interpretive statement and focus solely on the empirical test-set performance and the ensemble design choices.","revision_made":"yes","referee_comment":"[Abstract] Abstract: the central claim that 'the encoder's pretraining recipe, rather than its parameter count or the decoder design, is the dominant factor' is unsupported. The manuscript presents only a single ensemble configuration (all three encoders with Mask2Former under identical training/TTA) and its test mIoU; no tables or sections isolate pretraining effects while holding model size and decoder fixed, nor compare the three encoders individually under matched conditions."},{"response":"We acknowledge that no ablations on encoder removal or weighting variations are included. We will revise the method description to present the encoder selection (based on distinct pretraining objectives) and per-class weighting as design decisions motivated by the challenge setting, without claiming untested robustness or complementarity.","revision_made":"yes","referee_comment":"[Abstract / method description] The complementarity assumption underlying the ensemble (distinct pretraining objectives plus per-class IoU weighting) is stated but not tested for robustness. No ablation shows that removing any one encoder or altering the weighting scheme materially changes test performance, leaving open whether the reported gain is due to pretraining diversity or simply to ensembling three strong models."}],"tokens_in":1432,"tokens_out":347,"duration_ms":21154,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The punchline is that this is a competition technical report achieving second place on the GOOSE benchmark through an ensemble of three foundation vision encoders, but the assertion that pretraining dominates other factors lacks supporting experiments.\n\nThe authors pair DINOv3, SigLIP2, and InternImage with Mask2Former and apply a consistent training recipe including extended schedules, EMA, larger crops, and multi-scale flip TTA. They then weight the models per class based on validation IoU. This produces the reported 75.40% composite mIoU on the held-out test set.\n\nWhat stands out is the clear documentation of the pipeline and the practical success on a challenging 56-class outdoor segmentation task from multiple camera platforms. The per-class weighting is a reasonable tactic for balancing performance across categories.\n\nThe soft spot sits in the central claim. The report offers no controlled comparisons that isolate pretraining effects from model size or decoder choice. There are no results for the individual encoders under identical conditions, no parameter-matched models with different pretrainings, and no alternative decoders. The complementarity of the three pretraining objectives is assumed rather than measured. Without those isolations the dominance statement remains an assertion.\n\nThis paper will interest participants in the GOOSE challenge or similar fine-grained outdoor segmentation problems who want a working recipe. It offers little for readers focused on methodological innovation or rigorous factor analysis. I would not recommend sending it for peer review; the value is the competition result, not a research contribution that requires referee evaluation.","headline":"This competition report gets second place with a foundation model ensemble but does not back its claim about pretraining dominance with any isolating ablations.","tokens_in":2257,"tokens_out":382,"would_cite":false,"duration_ms":23344,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Encoder pretraining recipe dominates accuracy over model size or decoder in outdoor fine-grained segmentation.","keywords":["semantic segmentation","foundation models","ensemble learning","outdoor scenes","fine-grained categories","pretraining objectives","vision encoders","Mask2Former"],"falsifier":"A single encoder or a different weighting scheme matching or exceeding 75.40 percent mIoU on the unseen test set would show that the pretraining-diverse ensemble is not required.","tokens_in":2590,"feed_emoji":"🌄","tokens_out":689,"duration_ms":19159,"temperature":0.7,"pith_summary":"The report shows that an ensemble of three foundation vision encoders pretrained with distinct objectives can parse unstructured outdoor scenes into 56 fine-grained categories with high accuracy. By weighting the encoders per class using validation IoU and pairing them with a standard decoder plus strong training and test-time augmentation, the method reaches 75.40 percent mIoU on the official test set. The authors conclude that pretraining differences drive performance more than parameter count or decoder architecture. This matters for building robust perception systems where choosing complementary pretraining sources can substitute for simply scaling models.","feed_headline":"Pretraining recipe trumps size in outdoor segmentation","feed_subtitle":"DINOv3-SigLIP2-InternImage ensemble with per-class weighting reaches 75.4% mIoU on 56-class GOOSE scenes by showing pretraining as the main","key_machinery":"Pretraining-diverse ensemble of foundation vision encoders (DINOv3, SigLIP2, InternImage) combined by per-class validation-IoU weighting","core_discovery":"Our submission pairs foundation vision encoders including DINOv3, SigLIP2, and InternImage with a Mask2Former decoder, trains them using long schedules, exponential moving average, larger crop size, and multi-scale plus flip test-time augmentation, and combines the three via per-class validation-IoU weighting to exploit complementary pretraining objectives, achieving 75.40 percent composite mIoU and second place while demonstrating that the encoder pretraining recipe is the dominant accuracy factor rather than parameter count or decoder design.","pith_inferences":["The result suggests selecting future encoders by pretraining data diversity rather than architecture scale when facing unstructured outdoor scenes.","The same weighting approach could be tested on other segmentation benchmarks to check whether pretraining complementarity reduces the need for task-specific data.","If pretraining remains dominant, lighter decoders might suffice once the right encoder mix is chosen."],"forward_implications":["Encoders with distinct pretraining objectives cover fine-grained outdoor categories more effectively than any one encoder alone.","Long training schedules, EMA, larger crops, and multi-scale flip TTA reliably improve the ensemble output.","Pretraining choice can be prioritized over increasing model parameters when accuracy on this benchmark is the goal.","Per-class IoU weighting produces a combination that generalizes from validation to the official test distribution."],"fun_headline_variants":["Pretraining recipe key to accuracy on GOOSE scenes","Diverse encoder pretraining for GOOSE 2D segmentation","Pretraining dominant in foundation vision encoder ensemble","75.4% mIoU via pretraining diverse encoders on GOOSE"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"The three encoders supply enough complementary information from their different pretraining objectives that per-class validation-IoU weighting will combine them without overfitting to the validation set.","fun_headline_variants_meta":{"raw":{"variants":["Pretraining recipe key to accuracy on GOOSE scenes","Diverse encoder pretraining for GOOSE 2D segmentation","Pretraining dominant in foundation vision encoder ensemble","75.4% mIoU via pretraining diverse encoders on GOOSE"]},"model":"grok-4.3","cost_usd":0.006806,"raw_usage":{"total_tokens":3172,"prompt_tokens":684,"num_sources_used":0,"completion_tokens":66,"cost_in_usd_ticks":68062000,"prompt_tokens_details":{"text_tokens":684,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2422,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":684,"tokens_out":66,"duration_ms":10537,"temperature":1.0,"reasoning_tokens":2422,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-26T08:42:04.778342+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A single encoder or a different weighting scheme matching or exceeding 75.40 percent mIoU on the unseen test set would show that the pretraining-diverse ensemble is not required.","supporting_citations":[],"review_version":1}