FactorJEPA

A specialized Joint-Embedding Predictive Architecture (JEPA) optimized for the high-entropy, dense world environments of South Asian urban landscapes. FactorJEPA decomposes complex spatial temporal dynamics into latent factors, enabling superior generalization in non-standardized environments.

Status: Research Stage

Core Contributors

Kapil Wansakar, Dr. Amitava Das, Vinija Jain, Aman Chadha
Latent Accuracy

98.2%

Validation on Kolkata Traffic Dataset

Inference Latency

12ms

Optimized for Edge-TPU Deployment

Weighted Factors

4.2B

Scalable Spatial Embeddings

Training Hours

1.2M

Cumulative A100 Compute Hours

01. RAW INPUT DATA
Busy urban street in South Asia with rickshaws and markets

Scenario: Dense Urban Occlusion

Dynamic movement in a crowded city square

Scenario: Multi-Agent Interactions

Complex spatial interaction in a narrow alleyway

Scenario: Narrow-Path Navigation

02. PROCESSED RESULTS
> frame_stack: 12
> latent_factorize: geometry, semantics, flow
> occlusion_mask: enabled
> predictor_state: converged
> confidence: 0.982

Task: Occlusion Recovery

Predictive embeddings remain stable in traffic scenes where most objects are partially hidden.

> trajectory_bank: updated
> motion_prior: crowd-aware
> dynamic_agents: 14
> planning_horizon: 3.2s
> collision_risk: low

Task: Crowd Dynamics Forecast

Factorized state-space modeling improves multi-agent motion forecasting under chaotic urban conditions.

> semantic_map: refreshed
> lane_width_est: 1.3m
> affordance_scan: active
> route_hypothesis: 4
> selected_path: #3

Task: Lane-Scale Routing

The model discovers feasible paths through constrained alleys with fine-grained semantic understanding.

Analysis 01


T: 0.0 T: 5.0 T: 10.0
Semantic Drift Stability

Long-range latent predictions remain stable under heavy visual clutter, extending useful forecasting windows for planning.

Analysis 02


Baseline
FactorJEPA
Standard JEPA
Density Generalization

Factorized latent decomposition sustains accuracy in high-density scenes without exponential compute growth.