Dataset for OOF prediction - processes both halves separately but tracks original image. Returns: (image_tensor_left, image_tensor_r

EVA-CLIP-based model with 5 separate heads for all targets. Architecture: 1. EVA-CLIP Vision Encoder (768-dim embeddings for EVA02-C

Timm-based model with 5 separate heads for all targets. Architecture: 1. timm Vision Model (e.g., eva02_large_patch14_448.mim_m38m_f

Timm-based 5-head model with Height prediction and FiLM conditioning. Key innovation (kmat-style): - Predicts Height from image (auxilia

Dataset for OOF prediction - processes both halves separately but tracks original image. Returns: (image_tensor_left, image_tensor_r

Dataset for 3-crop OOF prediction. Returns all 3 crops for each image: - Left: 0-1000 - Center: 500-1500 - Right: 1000-2000

Training Dataset with image splitting and labels. Each 2000x1000 image is split into left and right halves. Labels are HALVED since each

EVA-CLIP-based model with gradual unfreezing capability. Architecture: 1. EVA-CLIP Vision Encoder (768-dim embeddings for EVA02-CLIP

EVA-CLIP model with 5 binning heads for biomass prediction. Each head outputs logits for bin classification. At inference, converts to e

InternViT-based model with 5 separate heads for all targets. Uses LOCAL InternViT implementation (no flash_attn dependency). Weights are

OpenCLIP-based model with 5 separate heads for all targets. Uses open_clip to load models from HuggingFace Hub. This is required for mod

Timm-based model with gradual unfreezing capability (3-head version). Architecture: 1. timm Vision Model (e.g., eva02_large_patch14_

Timm-based model with Residual Density Head. Architecture: 1. DINOv3 ViT backbone: - CLS token → Global prediction (stabl

Timm-based model with Gaussian output for uncertainty-aware predictions. Instead of predicting point estimates, this model outputs mean (mu)

Timm-based model with Soft-Species Conditioning. Architecture: 1. timm Vision Model → global pooled embedding (embed_dim) 2.

Pure Density Integration Model (No Global Head). Architecture: 1. DINOv3 ViT backbone extracts patch tokens (32×32 for 512px input)

Weighted SmoothL1Loss for 5-head model with consistency loss.

CrossEntropyLoss for binning classification. Converts continuous targets to bin indices and computes weighted CE loss. Args: r2

Dataset with Mosaic augmentation for biomass prediction. Mosaic combines 4 images into a 2x2 grid: - Each image is resized to half size

Dataset for OOF prediction with TTA support. Returns 4 views for each image: - Original - HorizontalFlip - VerticalFlip - HF

Dataset for OOF prediction with TTA support. Returns 4 views for each image: - Original - HorizontalFlip - VerticalFlip - HF

Dataset for EVA-CLIP-based biomass prediction with image splitting. Each 2000x1000 image is split into: - Left half: 1000x1000 - Rig

Dataset for biomass prediction with image splitting. Each 2000x1000 image is split into: - Left half: 1000x1000 - Right half: 1000x1

Dataset for 3-crop training with overlapping crops. Each 2000x1000 image is split into 3 overlapping crops: - Left: 0-1000 (50% of i

Dataset for sum-loss training (full-label supervision). Each sample returns BOTH halves together so training can compute: pred_full

Dataset for View Consistency training. For each sample, returns: - Original view (with random augmentation) - Flipped view (HFlip or

Weighted loss with consistency constraints for biomass prediction. Enforces physical constraints: 1. Dry_Dead_g + GDM_g ≈ Dry_Total_

Combined loss for 5-head biomass model with physical consistency constraints. Loss = SmoothL1Loss(predictions, targets) + consistency_weight

EVA-CLIP-based Model with Frozen Embeddings and Trainable Linear Heads. Must match EXP/EXP030/train.py architecture exactly.

EVA-CLIP-based model with 5 separate heads for all targets (inference version). Architecture: 1. EVA-CLIP Vision Encoder (768-dim em

EVA-CLIP-based model with 5 separate heads for all targets.

Gaussian NLL Loss with consistency constraints for biomass prediction. Instead of predicting point estimates, the model predicts mean (mu) a

Gradient Reversal Layer wrapper. Usage: grl = GradientReversalLayer(lambda_=0.1) reversed_features = grl(features)

Gradient Reversal Layer wrapper. Usage: grl = GradientReversalLayer(lambda_=0.1) reversed_features = grl(features)

Ranking Loss for learning relative ordering (A > B, A < B). Key insight: SigLIP/CLIP models use semantic shortcuts ("Clover present" → predi

Spatial-Deep Density Head for fast inference.

Batch sampler that ensures value diversity within each batch. For ranking loss to be effective, each batch should contain samples with d

Test Dataset for EVA-CLIP-based pipeline with image splitting. Each 2000x1000 image is split into: - Left half: (0, 0, 1000, 1000) -

Test Dataset for inference (no labels).

Test Dataset for timm-based pipeline with image splitting. Each 2000x1000 image is split into: - Left half: (0, 0, 1000, 1000) - Rig

Test Dataset with image splitting (No TTA).

Test Dataset for 3-crop inference (matches train.py BiomassDatasetOOF3Crop). Each 2000x1000 image is split into 3 overlapping crops: - L

Test Dataset with blur preprocessing (no TTA, single view). Rationale: - Train images have varying sharpness (Laplacian variance: 181-10

Test Dataset with blur-based TTA (no flip/rotate). Rationale: - Train images have varying sharpness (Laplacian variance: 181-10030, 55x

Test Dataset with crop-based TTA (physically consistent for this pipeline). Motivation: - Training uses RandomResizedCrop in some augmen

Test Dataset with photometric TTA (no flip/rotate). Rationale: - In this competition, the hidden test split differs by Sampling_Date, wh

Test Dataset with TTA (Test Time Augmentation) support. TTA applies 4 augmentations to each image: - Original - HorizontalFlip -

Test Dataset with TTA (Test Time Augmentation) support. TTA applies 4 augmentations to each image: - Original - HorizontalFlip -

Timm-based Model with Frozen Embeddings and Trainable Linear Heads. Must match the EXP060 timm training architecture (3-head).

Timm-based model with 5 separate heads for all targets (inference version).

Timm-based model with Residual Density Head (inference version). Architecture: 1. DINOv3 ViT backbone: - CLS token → Glob

Timm-based model with Gaussian output for uncertainty-aware predictions (inference version). This model outputs mean (mu) for each target. V

Timm-based 5-head model with Height prediction and FiLM conditioning (inference version). Key innovation (kmat-style): - Predicts Height

Timm-based model with Soft-Species Conditioning (inference version). Architecture: 1. timm Vision Model → global pooled embedding (e

Pure Density Integration Model (No Global Head) - Inference version. Architecture: 1. DINOv3 ViT backbone extracts patch tokens (32×

Pure Density Integration Model optimized for fast inference. - Supports FP16 inference - Minimal overhead

Weighted SmoothL1Loss for biomass prediction.

Validation Dataset - returns both halves for OOF prediction.

Gradient Reversal Layer (GRL) for Domain Adversarial Training. Forward pass: identity Backward pass: negate gradients and scale by lambd

Gradient Reversal Layer (GRL) for Domain Adversarial Training. Forward pass: identity Backward pass: negate gradients and scale by lambd

Residual density head: stable 1×1 baseline + small dilated 3×3 context correction. Starts identical to the 1×1 baseline by zero-init

Residual density head: stable 1×1 baseline + small dilated 3×3 context correction.

Spatial-Deep Density Head: DWConv for spatial context + MLP for non-linearity. Architecture: LayerNorm → 3x3 DWConv → Ba

Spatial-Deep Density Head: DWConv for spatial context + MLP for non-linearity. Inference version (no initialization needed - weights