Support for nuScenes dataset - #1
Open
sjsunadddy wants to merge 1 commit into
Open
Conversation
lkk688
added a commit
that referenced
this pull request
Jul 1, 2026
Document the #1 result from the GaussianFormer3D study: porting Lovasz-softmax + inverse-freq class weights onto the occ head roughly doubles mIoU at a fixed cheap setting (DINOv2-small/1k/8ep), 3-seed, no geo-IoU trade-off. Beats the 3x-bigger 0.216 run. Attribution: two knobs redundant not additive (same rare-class-starvation fix). TUTORIAL 2.8.1 + results table + README. Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
lkk688
added a commit
that referenced
this pull request
Jul 14, 2026
New doc: camera-only SOTA ~44.5 (COTR/SHTOcc, temporal) / fusion SOTA 56.6 (FusionOcc); our leaked cam 30.2 (CTF-Occ tier, ~15 below SOTA) / fusion 55.8 (leak inflates to ~SOTA, not real; clean ~45-50). Recent trends: temporal, 3D-Gaussian, sparse voxels, RayIoU, VGGT geometry-foundation backbones. VGGT direction: VGGT-Det is indoor; the on-point driving refs are VG3T/DVGT/DriveVGGT (CVPR'26, VGGT->occ). Fits our #1 weakness (dead camera path); differentiate via occ-pretraining-for-detection. Planned ablation: VGGT frozen backbone vs DINOv2 on camera-only occ. VGGT cloned at Others/VGGT-Det-CVPR2026. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
lkk688
added a commit
that referenced
this pull request
Jul 15, 2026
…84% of LiDAR) 50 official-val frames. Frozen VGGT depth-lift (scale-aligned) geo-IoU 0.140: +51% over Depth-Anything mono (0.093), 84% of LiDAR single-sweep oracle (0.167). Caveat: VGGT depth non-metric (scale ~18.8x) -- needs metric/extrinsic-baseline head. Validates the trained --backbone vggt build (ablation #2) as the camera-path fix. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
lkk688
added a commit
that referenced
this pull request
Jul 15, 2026
…cross 4 ablations VGGT features backbone 0.196 vs DINOv2 0.293 (-33%): geometry-specialized tokens are the wrong bias for the semantic occ head. Combined #1-4: VGGT's label-free geometry (84% LiDAR) helps in NO trained integration -- depth prior redundant with supervision (#2/#3), features worse than semantic backbone (#4). Mechanism: geometry FM helps only where geometry is the bottleneck (label-free/self-sup regime), not supervised. Finding A complete. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
lkk688
added a commit
that referenced
this pull request
Jul 19, 2026
…ensifier, staged plan) raycast.py: LiDAR-ray free-space carving -> per-voxel occupied/ray-free/UNKNOWN. Both teachers now supervise occupied (conf) + ray-verified free (free_w) and IGNORE unknown (weight 0) -- real free-space supervision + no penalty on unobserved voxels. Applied to BOTH teachers (keeps the gate a clean representation comparison). Student loss uses teacher weight directly. Validated: ~25% supervised (occ+ray-free), ~75% unknown ignored; gaussian 2.5x occupied filling. Doc: VGGT-as-teacher-densifier (narrative flip), full target arch, current-vs-ideal, staged plan. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
lkk688
added a commit
that referenced
this pull request
Jul 19, 2026
…n over-fills) Minimal gate: gaussian marginally > voxel10 (mIoU 0.0947 vs 0.0913, tail 0.0486 vs 0.0467). Ray-aware (#1) FLIPPED: voxel10 wins all (mIoU 0.0920, geo 0.360, tail 0.0335 vs gaussian 0.0852/0.280/0.0305). Mechanism: isotropic gaussian over-fills (30.9k vs 12.1k occ); ray-verified free conflicts with over-fill -> voxel geo jumps 0.294->0.360, gaussian flat. Minimal 'win' was over-fill recall artifact. Verdict: isotropic gaussian doesn't beat voxel fairly -> anisotropic surface-aligned Sigma is the make-or-break next test (don't go 4D if it doesn't win tail-IoU). Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
lkk688
added a commit
that referenced
this pull request
Jul 19, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
This PR has all changes required to support nuScenes dataset.