Bipin Saha

Technical Writeups

Longer-form field notes on autonomous-driving perception — the parts that do not fit in a paper: implementation roadmaps, failure taxonomies, and evaluation practice. Written for researchers and engineers building the stack.

  1. From 2D Perception to BEV Fusion: An Implementation Roadmap

    Nov 2025 · updated Aug 2026 · ~22 min · 22 refs

    A staged path from a working 2D detector to a multi-sensor bird's-eye-view stack — monocular 3D, push vs. pull view transforms, LiDAR BEV, and camera–LiDAR fusion — with the coordinate-frame, training-recipe, and deployment pitfalls that bite at each stage.

  2. Camera–LiDAR Fusion in Practice: Failure Modes and Design Patterns

    Jan 2026 · updated Aug 2026 · ~24 min · 22 refs

    Fusion beats either sensor alone on clean benchmarks, then fails in correlated, systematic ways: calibration drift, temporal misalignment, modality laziness, and asymmetric weather degradation. A taxonomy of what breaks and the architectural patterns that contain it.

  3. Self-Supervised Perception for Autonomous Driving: A Practical Overview

    Feb 2026 · updated Aug 2026 · ~28 min · 35 refs

    Fleet vehicles record far more than anyone can label. A survey of contrastive, masked-modeling, rendering, and cross-modal-distillation pretext tasks for LiDAR and multi-sensor data — what actually transfers, where it does not, and how to evaluate it honestly.

  4. Vision–Language Approaches for Vehicle Detection in Complex Scenarios

    Mar 2026 · updated Aug 2026 · ~26 min · 18 refs

    Closed-set detectors fail on the long tail: region-specific vehicle types, dense occlusion, adverse weather, distribution shift. How CNN, transformer, and vision–language detector families compare on these cases, and a discovery-then-distillation pipeline for shipping the result.

  5. How to Build Reproducible Perception Benchmarks for AV Research

    Apr 2026 · updated Aug 2026 · ~29 min · 18 refs

    Why 3D-perception leaderboard gains so often fail to reproduce — hidden backbones, test-time augmentation, split leakage, seed variance — and a concrete checklist for benchmark, metric, and split design that supports fair, compute-matched comparison.