Removing noise from partial 3D plant scans was associated with lower reconstruction error when the cleaned scans were passed to a model that also used the plant's previous growth stage, according to an arXiv preprint. On the real Pheno4D dataset, the best reported pairing, GCN with Adaptive Temporal PoinTr, produced a test Chamfer Distance, or CD, of 0.0066. The same study found that the strongest denoiser changed between real and synthetic data, while a different reconstruction score favored another model.
The results are best read as within-dataset comparisons rather than a universal ranking. The manuscript is an arXiv v1 preprint dated 28 Aug 2026. That caution is especially important for the real-data denoising test, whose reference geometry was SOR-based rather than exact ground truth.
How the test was built
The researchers evaluated a two-stage pipeline. A denoising model first cleaned a partial scan. A completion model then used the current scan and, in Adaptive Temporal PoinTr, the previous growth stage to recover occluded regions. The architecture encodes current and prior scans into proxy tokens, refines current features spatially and uses cross-attention to query prior features.
Four denoisers and three completion models were combined into 12 pipelines and compared in raw and denoised conditions. Pheno4D contributed 74 held-out scans for reconstruction and 60 stage-level observations for traits. The synthetic SynthCrop4D set contained 20 generated plant instances across six growth stages, yielding 120 paired clean and noisy point clouds. Its trait evaluation used three held-out plants and 15 stage-level observations.
Noise mattered after the models were combined
In the standalone denoising benchmark on 74 held-out Pheno4D scans, test CD values were nearly identical after 30 training epochs. They ranged from 0.0860 to 0.0863, a reported spread of 0.0003. The authors relate that similarity to the SOR-based reference used for the real scans.
The combined pipeline showed a larger difference. For Adaptive Temporal PoinTr, reported real-data test CD was 0.0180 in the raw condition and 0.0066 after denoising, a reported reduction of 63.3 percent. For Vanilla PoinTr, the corresponding values were 0.0191 and 0.0075, with a reported reduction of 60.7 percent.
Temporal PoinTr also had lower reported CD than Vanilla PoinTr across the real-data denoiser pairings. The ranges were 0.0066 to 0.0072 for the temporal model and 0.0075 to 0.0088 for Vanilla PoinTr. The manuscript reports a paired t-test with a p-value below 0.0001 after 15 bootstrap rounds. This remains an association reported within the experiment, not evidence that temporal conditioning caused the difference.
The best denoiser depended on the data. GCN plus Adaptive Temporal PoinTr was the best reported real-data pairing at 0.0066, while Mamba-DG plus Adaptive Temporal PoinTr was best on SynthCrop4D at 0.0061. The authors attribute the difference to different noise structures, and the synthetic result comes from one procedural Gaussian-noise regime.
CD was not the whole story
The F-Score results produced a less uniform ranking. On real data, Vanilla PoinTr with Mamba-DG had the highest reported test F-Score, 0.2080, while Temporal PoinTr with GCN reached 0.2075. On synthetic data, Vanilla outscored Temporal for the cited GCN pairing, 0.1828 versus 0.1821, and for Mamba-DG, 0.1823 versus 0.1482.
The paper says F-Score is sensitive to strict point overlap and temporal phase lag, which helps explain why its ranking did not match CD. Taken together, the findings describe a metric- and domain-specific result rather than a single overall winner.
Trait estimates improved in a real-data variant
Trait estimates followed the same divide between real and synthetic data. In the baseline comparison, mean absolute error was higher on the real Pheno4D observations than on the synthetic observations for plant height, canopy width, hull volume and surface area. The real-versus-synthetic errors were 0.277 versus 0.06 metres for height, 0.185 versus 0.096 metres for canopy width, 0.343 versus 0.021 cubic metres for hull volume, and 1.033 versus 0.229 square metres for surface area.
In a real-data V2 evaluation, the table reported lower mean absolute error for all four traits. Reported improvements were 71.5 percent for height, 37.8 percent for canopy width, 70.3 percent for hull volume and 81.9 percent for surface area. The corresponding trajectory correlations were 0.94, 0.97, 0.84 and 0.90. V2 was evaluated only in the real-data experiments described in the preprint.
For tomato height, the reported phase lag was +3.97 stages for the baseline and +2.0 stages for V2. The reported V2 growth-curve inflection point was at stage 3.5, compared with stage 1.5 for the ground-truth trajectory. The paper says transient single-stage canopy events were not consistently anticipated.
What the results leave open
The real denoising reference was SOR-filtered, while the synthetic comparison used one procedural Gaussian-noise regime. Poisson reconstruction produced no watertight meshes in the reported evaluation, so surface-area and leaf-area-index estimates were approximate proxies. Canopy width was also noisy in daily scans because leaves can reorient.
The paper reports computational scaling concerns for Langevin sampling at large point counts and says single-step inference was not systematically benchmarked. It cites a 149-scan Pheno4D training set as a limitation and warns that direct numerical comparisons with prior methods are invalid when datasets, sensors, occlusion protocols and metric conventions differ.
Source code, implementation details and pretrained weights are stated to be publicly available on GitHub. The paper also says Pheno4D is publicly available and SynthCrop4D can be reproduced with codebase scripts. The authors declare no known competing financial interests or personal relationships that could have influenced the work. The findings therefore remain within-study comparisons of reconstruction and derived trait metrics.
Paper data and sources
Original title: Denoising-Aware Temporal Point Cloud Completion for 3D Crop Architecture Recovery and Phenotypic Trait Extraction
Authors: Mrudul Mittal, Soumyashree Kar
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text