drive link

Signed-off-by: Pranav Doma <pranavreddy2327@gmail.com>

Author: pranavreddy23

Models/model_library/AutoDrive/README.md

@@ -1,63 +1,32 @@
-## AutoDrive — end-to-end longitudinal + lateral cues
+## AutoDrive 2.0 - temporal end-to-end distance and curvature estimation
 
-AutoDrive is a compact end-to-end network for **driver-assistance style** outputs from a front camera: it predicts **distance to the closest in-path object (CIPO)**, **road curvature** (related to steering demand), and a **binary CIPO presence** signal. It is designed to sit alongside classical pipelines (for example AutoSpeed + homography-based distance) and to consume the same style of **wide front view** inputs after a fixed geometric crop.
+For autonomous cruise and driver-assistance applications, it is important to estimate both the road-following demand and the lead-object distance from camera input. AutoDrive is a compact temporal model that predicts three outputs from consecutive front-view frames: **distance-to-CIPO**, **road curvature**, and **CIPO presence probability**.
 
-The network shares its backbone architecture with the **AutoSpeed** detector (same width/depth variant) and adds a small **temporal head** that fuses **previous** and **current** frame features. Training typically **warm-starts the backbone** from a trained AutoSpeed checkpoint, then learns the head (and optionally the full stack) on labelled sequences.
+AutoDrive processes a pair of frames (`t-1`, `t`) and uses a shared-feature temporal fusion head to capture short-term motion cues that a single-frame model can miss. It is designed for 2:1 input aspect ratio and is typically used with the same center-crop preprocessing used during training.
 
-### What AutoDrive outputs (inference)
+### Demo Video
+[Watch the demo video](<ADD_DRIVE_DEMO_LINK_HERE>)
 
-- **Normalized distance** `d_norm` in `[0, 1]`  
-  Mapped to meters as:  
-  `d = 150 × (1 - d_norm)`  
-  *(capped at 150 m)*
+## Get Started
 
-- **Curvature** in `1/m`  
-  *(Scaled internally during training; see `CURV_SCALE` in `Models/data_utils/load_data_auto_drive.py`)*
+To quickly try AutoDrive on your own data, please follow the steps in the [tutorial](tutorial.ipynb).  
+For best results, ensure your inference input follows the training preprocessing pipeline (50-degree center crop and 1024x512 resize).
 
-- **CIPO logit**  
-  Converted to probability using the sigmoid function  
-  A common validation threshold is **0.65** for class 1
+### Performance Results
 
-Input resolution after the standard ZOD / training preprocessing is **1024 × 512** (2:1), RGB, with **ImageNet** mean/std normalisation.
+AutoDrive is trained as a multi-task regression/classification model on sequence data with labels for curvature, CIPO distance, and CIPO presence.  
+Please add your official release metrics here once final evaluation is complete.
 
-### Demo / explainer
+## Model variants
 
-A packaged demo video link can be added here when the public release is ready. Until then, use the [tutorial](tutorial.ipynb) and the inference scripts under `Models/inference/`.
+AutoDrive currently uses one primary variant in this repository:
 
-## Get started
+**AutoDrive 2.0 model weights - 2:1 aspect ratio, 1024px by 512px input image**
+- [Link to Download Pytorch Model Weights *.pt](<https://drive.google.com/drive/u/1/folders/182h_9eBHroMCOfQHJiXVgrNq7zHx7Qws?dmr=1&ec=wgc-drive-hero-goto>)
 
-Follow the steps in **[tutorial.ipynb](tutorial.ipynb)**. Use the same **2:1** aspect ratio as training (1024 × 512) for best results.
+### Notes
 
-### Where the code lives
-
-| Topic | Location |
-|--------|-----------|
-| Network definition | `Models/model_components/autodrive/autodrive_network.py` |
-| Data + crop / labels | `Models/data_utils/load_data_auto_drive.py` |
-| Training entry | `Models/training/train_auto_drive.py` |
-| Example inference (video) | `Models/inference/autodrive_curvature_video.py` |
-| Benchmark vs classical | `Models/inference/classical_vs_autodrive_benchmark.py` |
-| Side-by-side comparison video | `Models/inference/comparison_video.py` |
-
-### Performance and datasets
-
-Training and evaluation are tied to your **ZOD** (or compatible) label layout: curvature, `distance_to_in_path_object`, and `cipo_detected` per frame. Report numbers in your release notes when you freeze a checkpoint; this README intentionally stays version-agnostic.
-
-## Model weights (release placeholder)
-
-**Planned public artifact:** `AutoDrive.pt` (or `AutoDrive.pth`) — single-file weights aligned with the release tag.
-
-Until the official link is published:
-
-- Use checkpoints produced by this repo, for example  
-  `{zod_root}/training/autodrive/run002/checkpoints/AutoDrive_best.pth`
-- The training script saves a dict with a `"model"` key containing `state_dict`-compatible weights.
-
-When you publish the final file, update this section with:
-
-- **PyTorch** — link to `AutoDrive.pt`
-- **Optional ONNX / TensorRT** — add links here if you export them
-
-## Model variant
-
-There is a single **AutoDrive** architecture in-tree: **1024 × 512** RGB, **two-frame** input (previous + current), backbone compatible with **AutoSpeed** `.pt` for weight transfer.
+- Training entry point: `Models/training/train_auto_drive.py`
+- Core network: `Models/model_components/autodrive/autodrive_network.py`
+- Data preprocessing and scaling: `Models/data_utils/load_data_auto_drive.py`
+- AutoSpeed backbone warm-start is supported through `--autospeed-ckpt`