AutoSapiens Dashboard

# RGB-D Modality Fusion Strategy

**Expected Designs:** 3

**Concept:** 
The Sapiens baseline uses 'early fusion' by modifying the first patch embedding layer for 4-channel input. This potentially dis...

# Kinematic Attention Masking

**Expected Designs:** 3

**Concept:** 
Human pose estimation models often predict physically impossible poses because the attention mechanism lacks explicit structural c...

# Curriculum-Based Loss Weighting

**Expected Designs:** 2

**Concept:** 
Absolute depth is hard and creates noisy gradients early on. It needs a learned or structured loss weighting over epochs to st...

# Layer-Wise Learning Rate Decay & Progressive Unfreezing

**Expected Designs:** 6

**Concept:** 
Use structure in how the backbone learns to prevent catastrophic forgetting of the ViT's rich pre-trai...

# Depth-Aware Positional Embeddings

**Expected Designs:** 3

**Concept:** 
Provide explicit geometric awareness to the Transformer, which currently only uses standard 2D positional encodings. Inject ...

# idea006 — Training Data Augmentation for Generalization

**Expected Designs:** 6

**Baseline starting point:** `baseline.py` (equivalently `baseline/train.py`, `baseline/model.py`, `baseline/transfo...

# Depth-Bucket Positional Embeddings with Layer-Wise Fine-Tuning

**Expected Designs:** 3

## Starting Point

The baseline starting point for this idea is:

`runs/idea005/design001/code/`

That design...

# Continuous Depth Positional Encoding

**Expected Designs:** 4

## Starting Point

The baseline starting point for this idea is:

`runs/idea005/design001/code/`

That design delivered the best comple...

# Head Architecture Refinement: Query Initialization, Depth of Decoder, and Output Normalization

**Expected Designs:** 5

## Starting Point

The baseline starting point for this idea is:

`runs/idea0...

# idea010 — Multi-Scale Backbone Feature Aggregation

**Expected Designs:** 5

## Starting Point

The baseline starting point for this idea is:

`runs/idea004/design002/train.py`

That design produced...

# idea011 — LLRD Schedule Combined with Continuous Depth PE

**Expected Designs:** 4

## Starting Point

The baseline starting point for this idea is:

`runs/idea008/design003/code/`

That design prod...

# idea012 — Regularization for Generalization

**Expected Designs:** 5

## Starting Point

The baseline starting point for this idea is:

`runs/idea004/design002/train.py`

That design produced the be...

# idea013 — Joint Prediction Loss Reformulation

**Expected Designs:** 4

## Starting Point

The baseline starting point for this idea is:

`runs/idea004/design002/train.py`

That design produced the ...

# idea014 -- Best-of-Breed Combination: LLRD + Depth PE + Wide Head

**Expected Designs:** 3

## Starting Point

The baseline starting point for this idea is:

`runs/idea008/design003/code/`

That des...

# idea015 -- Iterative Refinement Decoder on SOTA Triple Combo

**Expected Designs:** 4

## Starting Point

The baseline starting point for this idea is:

`runs/idea014/design003/code/`

That design i...

# idea016 -- 2.5D Heatmap Representation with Soft-Argmax Decoding

**Expected Designs:** 4

## Starting Point

The baseline starting point for this idea is:

`runs/idea014/design003/code/`

Current S...

# idea017 -- Temporal Context via Adjacent-Frame Fusion

**Expected Designs:** 4

## Starting Point

The baseline starting point for this idea is:

`runs/idea014/design003/code/`

Current SOTA: val_mp...

# idea018 — Weight Averaging: EMA and SWA on the SOTA Triple-Combo

**Expected Designs:** 4

## Baseline starting point
`runs/idea014/design003/code/` (current SOTA: 106.85 mm val_mpjpe_body, 96.73 mm...

# idea019 -- Anatomical Structure Priors for Iterative Refinement

**Expected Designs:** 5

## Starting Point

The baseline starting point for this idea is:

`runs/idea015/design001/code/`

That desig...

# idea020 -- Refinement-Specific Loss and Gradient Strategy

**Expected Designs:** 5

## Starting Point

The baseline starting point for this idea is:

`runs/idea015/design004/code/`

That design is t...

# idea021 -- Anatomical Priors on the Two-Decoder SOTA

**Expected Designs:** 4

## Starting Point

The baseline starting point for this idea is:

`runs/idea015/design004/code/`

That design is the cu...