--- language: en tags: - lidar - object-detection - education --- # CS 498 homework data `hw3_data.zip` is the frozen, required HW3 Car-detection package (SHA-256 `d70aa8544b52d56ca0569110434562ff74296ee80f63c8cd8b1c47364fc79523`). It contains `train.npz`, `val.npz`, and `test.npz` over 5,977/730/774 drive-disjoint KITTI frames. The processed targets use `grid-aligned-v2`. `hw3_multiclass_data.zip` is a small **optional bonus-only** companion (SHA-256 `58811f16591d5ae9de97da8fa3bf871bfb65d101f8b6b5ac87666a5aa56e489c`). Its three aligned label archives add Pedestrian and Cyclist boxes/class IDs; its Car boxes exactly match the core package. Use the handout's `data/download_multiclass_data.py` after `data/download_data.py`. The companion does not replace core BEVs, split membership, or the required result protocol. Both packages derive from the [KITTI 3D object-detection training set](https://www.cvlibs.net/datasets/kitti/eval_object.php?obj_benchmark=3d). Use is educational in CS 498; cite KITTI. Validation controls all model and threshold decisions, and the final test split must remain held out until those choices are frozen. File checksums are in `SHA256SUMS`. ## HW4: compact NeRF Lego data `hw4_data.zip` contains the 200 x 200 Blender-format Lego subset used by the CS 498 HW4 handout: 50 training, 5 validation, and 12 held-out test poses. Its SHA-256 is `502c9216170111bd9a6b7f7204442b23b69020924a11229ec0a820f825b3fbad`. The handout downloader verifies every member against `hw4_SHA256SUMS` before installation. The images are derived from the NeRF synthetic Lego example data and retain the original camera transforms; use is limited to the course and other permitted educational use. ## Final Project - Track 1: Multi-View 3D Semantic Occupancy Prediction Release: Wednesday, November 4, 2026. Due: December 9, 2026, 23:59 America/Chicago. Use the current Track 1 student package from the course. `final-track1/data/manifest.json` indexes 220 per-house archives: 200 Train, 10 Val and 10 Test houses. Each independent sample has three overlapping RGB images from one camera center, optional depth, intrinsics and poses. Train/Val include prepared 25 cm targets; Test includes inputs only. Test ground truth is not public. No simulator installation or data generation is needed. From the released project root, use `python run.py download --split train` (or `--split val` / `--split test`). The downloader pins revision `192c75d954d62969db82e2e3a018de2f71d4a48b`, verifies SHA-256 checksums and safely extracts the requested split. The canonical prediction grid is `(40,14,80)` in `(z,y,x)` order. Ranking is completion mIoU macro-averaged across all ten Test houses. See the course handout for the target policy and allowed inputs. The supported execution paths are Ubuntu with an NVIDIA GPU and Google Colab. Colab is the fallback and is recommended for Mac and Windows users. Follow `SETUP.md` and `notebooks/track1_colab.ipynb` in the student package; run `python run.py doctor` before training. Public downloads need no HF account/token. Data are rendered using ProcTHOR and AI2-THOR; cite their papers. Educational redistribution is authorized for this course. Other homework packages above are unchanged. ## Final Project - Track 2: Complete 3D Scene Reconstruction 30 Dev / 10 Val / 20 Test posed-RGB scenes. Dev/Val include evaluation GT; Test includes RGB and calibration only. No scene generation is required. See [the Track 2 data guide](final-track2/README.md) for the pinned downloader, permitted use, local Val CSV rehearsal and lightweight Kaggle workflow. The student package is distributed by the course. Test targets remain private.