submission name:

av1_crf31_bicubic

upload zipped archive.zip

archive.zip

report.txt

=== Evaluation results over 600 samples ===
  Average PoseNet Distortion: 0.07131552
  Average SegNet Distortion: 0.00520704
  Submission file size: 1,002,966 bytes
  Original uncompressed size: 37,545,489 bytes
  Compression Rate: 0.02671336
  Final score: 100*segnet_dist + √(10*posenet_dist) + 25*rate = 2.03

does your submission require gpu for evaluation (inflation)?

no

did you include the compression script? and want it to be merged?

yes

additional comments

SVT-AV1 preset 0 at CRF 31, with corridor-aware preprocessing and a
sharpening pass at inflate.

Pipeline:

• Preprocess (corridor ROI denoise). PoseNet and SegNet both downsample
  to 512x384 and only care about the lower-center driving corridor.
  Everything outside (sky, buildings, hood, passing trees) is high-entropy
  content the encoder would otherwise spend bits on for no scoring benefit.
  The preprocess step draws a per-segment corridor polygon, hand-tuned on
  the four 300-frame chunks of 0.mkv since the mount drifts, feathers the
  mask by 24px, and on the outside region runs a YUV-space luma gaussian
  blur (strength 2.5) plus medium chroma average-pooling. Inside the
  corridor the frame is untouched. The output is written as an FFV1
  lossless intermediate so the encoder only sees the denoised frames.
• Encode. 45% Lanczos downscale, then SVT-AV1 preset 0, CRF 31,
  film-grain 22, keyint 180, scd=0. The score has a sqrt on the PoseNet
  term and a linear 25x weight on rate, so spending a few extra bits at
  CRF 31 (vs a tighter CRF) gives PoseNet headroom that more than pays
  for the rate hit at this operating point. Going lower than 31 flipped
  the balance the other way in my sweeps.
• Inflate (bicubic + unsharp). Decode, bicubic upscale back to the
  original resolution, then a 9-tap binomial unsharp mask at strength 0.27.
  Bicubic was a small but consistent PoseNet improvement over lanczos on
  this content; SegNet was unchanged. The unsharp recovers the
  high-frequency edges SegNet relies on without visible ringing on lane
  markings.
• CPU-only inflation. No GPU required. The inflate script uses CUDA or
  MPS if available, otherwise falls back to CPU with identical output.

Final local score: 2.03 (posenet 0.0713, segnet 0.00521, rate 0.0267).
Evaluated on an M-series Mac (MPS). On the same box baseline_fast measured
4.43 vs the README's 4.39, so the CI number should land at or just under
2.03.