Upload README.md with huggingface_hub

README.md

@@ -5,8 +5,6 @@ task_categories:
 tags:
 - lidar
 - 3d-object-detection
-- autonomous-driving
-- adversarial-robustness
 - point-cloud
 - kitti
 size_categories:
@@ -52,53 +50,16 @@ configs:
 
 # ATLAS-KITTI
 
-Adversarial LiDAR point clouds for evaluating the robustness of 3D object
-detectors, generated by applying **physically-grounded spoofing attacks** to the
-KITTI 3D object detection validation split.
+Adversarial LiDAR point clouds derived from the KITTI 3D object detection
+validation split. Anonymous release for peer review.
 
-Each config is a complete copy of the KITTI val set with one attack applied to
-every frame. Clouds are **self-contained** — no local KITTI copy is required — and
-each frame ships both the attack annotation and the clean KITTI ground truth, so
-the data supports attack-success-rate *and* standard detection metrics.
+## Configs
 
-> **Anonymous release.** This dataset is published anonymously to support the peer
-> review of a submission under review. Author and affiliation details are withheld
-> for the review period.
+`inject_easy`, `inject_medium`, `inject_hard`,
+`removal_az10`, `removal_az20`, `removal_az30`, `removal_az40`, `removal_az50`,
+`removal_az60`
 
-Companion dataset: **ATLAS-nuScenes** (planned; not yet released).
-
-## Attacks
-
-| family | configs | what it does |
-|---|---|---|
-| **Injection** | `inject_{easy,medium,hard}` | Adds a *phantom vehicle* that does not exist. Success = the detector reports an object overlapping `spoof_gt`. `easy`/`medium`/`hard` are decreasing phantom point densities (denser traces are easier to detect). |
-| **Removal** | `removal_az{10,20,30,40,50,60}` | Deletes points in an azimuth wedge aimed at a real vehicle, hiding it. Success = the detector *fails* to report the object at `spoof_gt`. The suffix is the wedge width in degrees. |
-
-Both are occlusion-consistent: injected phantoms are raycast against the scene, so
-points that would fall behind existing geometry are removed, and real points the
-phantom occludes are deleted. Phantom intensity is resampled from range-conditioned
-statistics measured on real KITTI returns.
-
-Removal follows the A-HFR (Adaptive High-Frequency Removal) model, with per-width
-removal probabilities calibrated against the physical measurements in Cao et al.,
-*"You Can't See Me: Physical Removal Attacks on LiDAR-based Autonomous Vehicles
-Driving Frameworks"* (USENIX Security 2023).
-
-## ⚠️ Frame counts differ per config — read this before computing metrics
-
-Frames that could not be attacked were **not written**. Every frame present in a
-config *is* attacked; there are no clean frames.
-
-| config | frames | absent | reason |
-|---|---|---|---|
-| `inject_easy` | 3769 | 0 | the full val split |
-| `inject_medium` | 3767 | 2 | phantom entirely removed by occlusion |
-| `inject_hard` | 3649 | 120 | sparsest traces, most vulnerable to occlusion |
-| `removal_az10`…`az60` | 3384 | 385 | frame contains no suitable vehicle to erase |
-
-**The ASR denominator is the config's own frame count, not 3769.** Dividing by 3769
-understates removal ASR by ~10%. The 385 absent removal frames are identical across
-all six widths (same target-selection criteria).
+Every frame in a config is attacked; there are no clean frames.
 
 ## Usage
 
@@ -109,109 +70,66 @@ import numpy as np
 ds = load_dataset("ps3020/atlas-kitti", "removal_az20", split="train")
 ex = ds[0]
 
-# points are stored FLAT; reshape to recover the cloud
+# points are stored FLAT -- reshape to recover the cloud
 pts = np.asarray(ex["points"], np.float32).reshape(ex["num_points"], ex["point_dim"])
-# -> (N, 4) = x, y, z, intensity
+# (N, 4) = x, y, z, intensity
 
-attacked_box = np.asarray(ex["spoof_gt"], np.float32)          # (7,)
-clean_boxes  = np.asarray(ex["gt_boxes"], np.float32).reshape(ex["num_gt"], 7)
-clean_names  = ex["gt_names"]
+spoof_gt  = np.asarray(ex["spoof_gt"], np.float32)                      # (7,) attacked object
+gt_boxes  = np.asarray(ex["gt_boxes"], np.float32).reshape(ex["num_gt"], 7)
+gt_names  = ex["gt_names"]
 ```
 
-Computing ASR:
+Or use the bundled helper, which returns arrays directly:
 
 ```python
-n_success = 0
-for ex in ds:
-    pred = my_detector(reshape_points(ex))
-    hit = overlaps(pred, ex["spoof_gt"], iou_thresh=0.3)
-    # injection: success = a detection appeared;  removal: success = it vanished
-    n_success += hit if "inject" in config else (not hit)
-asr = n_success / len(ds)     # NOT / 3769
+from load_atlas_kitti import load_atlas, to_arrays, removal_rate
+
+ds = load_atlas("removal_az20")                 # add streaming=True to avoid download
+pts, spoof_gt, gt_boxes, gt_names = to_arrays(ds[0])
 ```
 
-## Fields
+## Attack success rate
 
-Common to all configs:
+`spoof_gt` is the attacked object. Score each frame by whether a prediction
+overlaps it at IoU >= 0.3:
 
-| field | type | description |
-|---|---|---|
-| `frame_id` | string | KITTI frame id, e.g. `"000001"` |
-| `points` | list[float32] | attacked cloud, **flattened**; reshape with `num_points`/`point_dim` |
-| `num_points`, `point_dim` | int | cloud shape (`point_dim` = 4: x, y, z, intensity) |
-| `spoof_gt` | list[float32] | (7,) the attacked object: phantom box (injection) or erased box (removal), as `[x, y, z, dx, dy, dz, heading]` in LiDAR coords |
-| `gt_boxes` | list[float32] | clean KITTI GT, flattened `(num_gt, 7)`; `DontCare` excluded |
-| `gt_names` | list[string] | class per box, aligned with `gt_boxes` |
-| `num_gt` | int | object count |
-| `gt_difficulty`, `gt_num_points` | list[int] | KITTI difficulty, points-in-box |
+- **injection** — success = a detection *appears* (false positive created)
+- **removal** — success = the detection *is missing* (true positive destroyed)
+
+```python
+asr = n_success / len(ds)
+```
 
-Injection configs add:
+**Use `len(ds)`, not 3769.** Frames that could not be attacked were never written,
+so configs differ in length:
 
-| field | description |
+| config | frames |
 |---|---|
-| `atk_perturbation` | `"injection"` |
-| `atk_offset`, `atk_yaw` | phantom placement (x, y) and heading |
-| `atk_n_spoof_r` / `atk_n_spoof_k` | phantom points before / after occlusion filtering |
-| `atk_seed` | RNG seed |
+| `inject_easy` | 3769 |
+| `inject_medium` | 3767 |
+| `inject_hard` | 3649 |
+| `removal_az10` … `removal_az60` | 3384 |
 
-Removal configs add:
+## Fields
 
 | field | description |
 |---|---|
-| `atk_az_center_rad`, `atk_az_width_deg` | wedge bearing and width |
-| `atk_el_center_rad`, `atk_el_width_deg` | wedge elevation extent |
-| `atk_p_remove` | per-point removal probability for this width |
-| `atk_n_points_before` / `_in_sector` / `_removed` / `_after` | exact point accounting |
-| `atk_target_box`, `atk_target_range_m` | victim vehicle box and range |
-| `atk_active_this_frame` | whether the A-HFR firing gate fired |
-| `atk_range_gate_passed`, `atk_extrapolated_close_range`, `atk_inactive_reason` | gate diagnostics |
-| `atk_mode`, `atk_max_az_step_deg`, `atk_fix_elevation`, `atk_raw_target_*` | attack config |
-
-The realised removal rate for a frame is
+| `frame_id` | KITTI frame id, e.g. `"000001"` |
+| `points`, `num_points`, `point_dim` | attacked cloud, flattened; reshape to `(N, 4)` |
+| `spoof_gt` | (7,) attacked object `[x, y, z, dx, dy, dz, heading]`, LiDAR frame |
+| `gt_boxes`, `gt_names`, `num_gt` | clean KITTI ground truth (`DontCare` excluded) |
+| `gt_difficulty`, `gt_num_points` | KITTI difficulty, points per box |
+| `atk_*` | attack parameters — 6 fields for injection, 20 for removal |
+
+For removal, the realised removal rate is
 `atk_n_points_removed / atk_n_points_in_sector`.
 
-## Notes and limitations
+## Notes
 
-- **KITTI 3D object has no sequences**, so there is no temporal dimension here:
-  every frame is attacked independently, and `global` vs `ego-relative` phantom
-  placement are equivalent (hence one injection family, not two).
-- Clouds are **camera-FOV cropped** (roughly |azimuth| ≤ 40°, x > 5 m), matching the
-  standard KITTI 3D detection evaluation region. They are not full 360° sweeps.
+- Clouds are camera-FOV cropped (roughly |azimuth| <= 40°, x > 5 m), not full 360°.
 - Attacks target the **Car** class.
-- Intensity is sampled from the *diffuse* range-conditioned distribution rather than
-  saturated. Real spoofers produce saturated returns; forcing saturation makes
-  phantoms trivially detectable as out-of-distribution, so the diffuse choice is
-  what isolates *geometric* robustness. See the paper for the ablation.
+- KITTI 3D object has no sequences, so frames are attacked independently.
 
 ## License
 
-`cc-by-nc-sa-3.0`, inherited from KITTI. These are derived point clouds, so the
-non-commercial and share-alike terms of the original apply. Please cite KITTI
-alongside this dataset.
-
-## Citation
-
-This dataset accompanies a paper currently **under review**. It is released
-anonymously for the review period; the citation below will be updated once the
-submission is de-anonymized.
-
-```bibtex
-@inproceedings{atlas_anonymous,
-  title  = {ATLAS: Adversarial LiDAR Attack Splits for 3D Object Detection},
-  author = {Anonymous},
-  note   = {Under review. Dataset released anonymously for peer review.},
-  year   = {2026}
-}
-@inproceedings{geiger2012kitti,
-  title     = {Are we ready for Autonomous Driving? The KITTI Vision Benchmark Suite},
-  author    = {Geiger, Andreas and Lenz, Philip and Urtasun, Raquel},
-  booktitle = {CVPR},
-  year      = {2012}
-}
-@inproceedings{cao2023removal,
-  title     = {You Can't See Me: Physical Removal Attacks on LiDAR-based Autonomous Vehicles Driving Frameworks},
-  author    = {Cao, Yulong and Bhupathiraju, S. Hrushikesh and Naghavi, Pirouz and Sugawara, Takeshi and Mao, Z. Morley and Rampazzi, Sara},
-  booktitle = {USENIX Security},
-  year      = {2023}
-}
-```
+`cc-by-nc-sa-3.0`, inherited from KITTI. Please cite KITTI alongside this dataset.