This document details our winning methodology for the CSIRO - Image2Biomass competition. Our solution addresses the complex regression task of quantifying pasture biomass from high-variance imagery. By leveraging cutting-edge Foundation Models (DinoV3, DinoV2) and implementing a robust ensemble strategy, we overcame significant challenges related to severe illumination shifts, occlusion, and extreme data scarcity. Our final architecture integrates Multi-Scale Patch Learning, Multiple Instance Learning (MIL), and advanced feature aggregation techniques to achieve superior generalization.
\nThe core challenge involved regressing biomass values ($kg/ha$) from unstructured visual data under unconstrained environmental conditions.
\nOur strategy shifted from training scratch architectures to leveraging the semantic richness of Self-Supervised Vision Transformers. We constructed a heterogeneous ensemble of 5 distinct regression pipelines, stabilizing the aleatoric uncertainty inherent in the data.
\nTo satisfy computational constraints while maximizing throughput, we architected a distributed inference pipeline, parallelizing the massive feature extraction workload across dual GPU nodes.
\nWe conducted extensive experiments to validate our hypothesis space. Due to the low-data regime, classical deep learning techniques proved detrimental:
\nOur ensemble demonstrates a significant performance uplift over individual weak learners.
\n| Model Architecture | \nPublic Score (R2) | \n
|---|---|
| I. DinoV3 (2x2 Patching) | \n0.74 | \n
| II. DinoV3 (3x3 Patching) | \n0.74 | \n
| III. MIL Transformer (DinoV2) | \n0.74 | \n
| IV. DinoV3 (Dense Stride) | \n0.76 | \n
| V. DinoV3 (Large Patch) | \n0.75 | \n
| Final Ensemble (Weighted Blend) | \n0.779 | \n
This document details our winning methodology for the CSIRO - Image2Biomass competition. Our solution addresses the complex regression task of quantifying pasture biomass from high-variance imagery. By leveraging cutting-edge Foundation Models (DinoV3, DinoV2) and implementing a robust ensemble strategy, we overcame significant challenges related to severe illumination shifts, occlusion, and extreme data scarcity. Our final architecture integrates Multi-Scale Patch Learning, Multiple Instance Learning (MIL), and advanced feature aggregation techniques to achieve superior generalization.
\nThe core challenge involved regressing biomass values ($kg/ha$) from unstructured visual data under unconstrained environmental conditions.
\nOur strategy shifted from training scratch architectures to leveraging the semantic richness of Self-Supervised Vision Transformers. We constructed a heterogeneous ensemble of 5 distinct regression pipelines, stabilizing the aleatoric uncertainty inherent in the data.
\nTo satisfy computational constraints while maximizing throughput, we architected a distributed inference pipeline, parallelizing the massive feature extraction workload across dual GPU nodes.
\nWe conducted extensive experiments to validate our hypothesis space. Due to the low-data regime, classical deep learning techniques proved detrimental:
\nOur ensemble demonstrates a significant performance uplift over individual weak learners.
\n| Model Architecture | \nPublic Score (R2) | \n
|---|---|
| I. DinoV3 (2x2 Patching) | \n0.74 | \n
| II. DinoV3 (3x3 Patching) | \n0.74 | \n
| III. MIL Transformer (DinoV2) | \n0.74 | \n
| IV. DinoV3 (Dense Stride) | \n0.76 | \n
| V. DinoV3 (Large Patch) | \n0.75 | \n
| Final Ensemble (Weighted Blend) | \n0.779 | \n
Sorry, there is a mistake in table. It's R2 score, not RMSE loss
", "rawMarkdown": "Sorry, there is a mistake in table. It's R2 score, not RMSE loss", "votes": null }, { "id": "3402217", "postDate": "02/05/2026 13:30:08", "content": "amazing job 🤯
", "rawMarkdown": "amazing job 🤯", "votes": null }, { "id": "3402218", "postDate": "02/05/2026 13:35:57", "content": "Thank you sir. I really appreciate it
", "rawMarkdown": "Thank you sir. I really appreciate it", "votes": null } ], "comments": [ { "id": 3402209, "author_name": "artemgoncarov", "author_url": "", "post_date": "02/05/2026 13:07:36", "content": "Sorry, there is a mistake in table. It's R2 score, not RMSE loss
", "votes": null, "replies": [] }, { "id": 3402217, "author_name": "d1ffic00lt", "author_url": "", "post_date": "02/05/2026 13:30:08", "content": "amazing job 🤯
", "votes": null, "replies": [ { "id": 3402218, "author_name": "artemgoncarov", "author_url": "", "post_date": "02/05/2026 13:35:57", "content": "Thank you sir. I really appreciate it
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3402202": "# CSIRO - Image2Biomass: State-of-the-Art Biomass Quantification Framework\n\n## Abstract\nThis document details our winning methodology for the CSIRO - Image2Biomass competition. Our solution addresses the complex regression task of quantifying pasture biomass from high-variance imagery. By leveraging cutting-edge Foundation Models (DinoV3, DinoV2) and implementing a robust ensemble strategy, we overcame significant challenges related to severe illumination shifts, occlusion, and extreme data scarcity. Our final architecture integrates Multi-Scale Patch Learning, Multiple Instance Learning (MIL), and advanced feature aggregation techniques to achieve superior generalization.\n\n## Problem Formulation & Challenges\nThe core challenge involved regressing biomass values ($kg/ha$) from unstructured visual data under unconstrained environmental conditions.\n* **High-Variance Illumination:** Drastic dynamic range differences complicated the semantic segmentation of senescent (dry) vs. photosynthetic (green) vegetation.\n* **Small-Shot Regime:** The limited dataset size posed a high risk of overfitting, rendering traditional fully-supervised training of large CNNs or ViTs ineffective.\n\n## Methodology: Ensemble of Foundation Models\nOur strategy shifted from training scratch architectures to leveraging the semantic richness of Self-Supervised Vision Transformers. We constructed a heterogeneous ensemble of 5 distinct regression pipelines, stabilizing the aleatoric uncertainty inherent in the data.\n\n### Architecture I: Coarse-Grained Spatial Aggregation (2x2 Grid)\n* **Backbone:** **Dino V3 ViT-7B** (State-of-the-Art Foundation Model).\n* **Feature Extraction:** Utilized a 4-patch spatial decomposition (2x2 Grid) to capture local-global context.\n* **Regression Head:** A lightweight Multi-Layer Perceptron (MLP) trained on frozen embeddings.\n* **Loss Function:** **SmoothL1 Loss** was selected over MSE to reduce sensitivity to label outliers and improve convergence stability.\n* **Robustness:** Implemented Seed Averaging (Stochastic Weight Averaging proxy) to minimize variance.\n* **Data Cleaning:** Manually curated the dataset to exclude samples with ground-truth label noise (2 instances removed).\n\n### Architecture II: Fine-Grained Spatial Aggregation (3x3 Grid)\n* **differentiation:** Increased spatial resolution of feature extraction.\n* **Preprocessing:** Image decomposition into a 9-patch grid (3x3), allowing the model to focus on finer textural details of the pasture structure.\n* **Backbone & Head:** Identical MLP projection on Dino V3 ViT-7B embeddings.\n\n### Architecture III: Deep Multiple Instance Learning (MIL)\n* **Architecture:** **MilTransformer**. Treated the biomass estimation as a bag-of-features problem.\n* **Backbone:** **Dino V2 ViT-G14 (Register-Equipped)**. The register tokens enhance the model's ability to discard background noise.\n* **Optimization:** Trained via standard MSE Loss.\n* **Augmentation Strategy:** Introduced Vertical Flips to enforce rotation invariance without distorting the biomass distribution prior.\n\n### Architecture IV: Dense Strided Feature Extraction\n* **differentiation:** High-density feature sampling.\n* **Mechanism:** Employed a sliding window approach with a dense stride of 112px over Dino V3 ViT-7B embeddings. This ensures maximum coverage and overlap of critical visual features.\n\n### Architecture V: Large-Scale Receptive Field Analysis\n* **differentiation:** Expanded context window.\n* **Mechanism:** Increased patch size to 336px to capture macro-level biomass patterns and texture density, complementing the fine-grained models.\n\n## Engineering: Scaled Inference Pipeline\nTo satisfy computational constraints while maximizing throughput, we architected a distributed inference pipeline, parallelizing the massive feature extraction workload across dual GPU nodes.\n\n## Ablation Studies & Research Limitations\nWe conducted extensive experiments to validate our hypothesis space. Due to the low-data regime, classical deep learning techniques proved detrimental:\n* **Negative Results:** Standard CNN fine-tuning, aggressive TTA (Test Time Augmentation), and end-to-end Transformer training led to rapid overfitting.\n* **Synthetic Domain Adaptation:** We generated synthetic samples using **Nano Banana** generative models to augment the training distribution. While this degraded Public Leaderboard performance (likely due to distribution shift), post-hoc analysis revealed it yielded the **SOTA score on the Private Leaderboard**, suggesting improved generalization capability that was penalized by the public test split distribution.\n\n## Results Summary\nOur ensemble demonstrates a significant performance uplift over individual weak learners.\n\n| Model Architecture | Public Score (R2) |\n| :--- | :--- |\n| I. DinoV3 (2x2 Patching) | 0.74 |\n| II. DinoV3 (3x3 Patching) | 0.74 |\n| III. MIL Transformer (DinoV2) | 0.74 |\n| IV. DinoV3 (Dense Stride) | 0.76 |\n| V. DinoV3 (Large Patch) | 0.75 |\n| **Final Ensemble (Weighted Blend)** | **0.779** |", "3402209": "Sorry, there is a mistake in table. It's R2 score, not RMSE loss", "3402217": "amazing job 🤯", "3402218": "Thank you sir. I really appreciate it" }, "source": "meta" }