Summary

This PR releases an INT8 PTQ version of J6Gen2_Base/2.6.1, delivering significant latency reduction and throughput improvement with minimal impact on detection accuracy.

• Key performance gains:
  • Average end-to-end latency improved by 25.1% (48.9 ms -> 36.6 ms).
  • Throughput increased from 20.5 FPS to 27.3 FPS.
  • Backbone inference latency was reduced by up to 39.7%.
• Accuracy impact remained negligible:
  • Absolute mAP drop stayed within 0.25%.
• Overall:
  • INT8 quantization effectively accelerates compute-heavy backbone inference while preserving detection accuracy.

Please check more details in T4 internal link

Evaluation

mAP

• Class mAP for BEV Center Distance: 0.5m, 1.0m, 2.0m, 4.0m

  Eval Range: 0.0 - 50.0m

  Model version	mAP	car (64,520)	truck (6,947)	bus (2,275)	bicycle (1,379)	pedestrian (19,421)
  CenterPoint J6Gen2_Base/2.6.1_quant	84.50	94.70	81.75	91.68	80.14	74.24
  CenterPoint J6Gen2_Base/2.6.1	84.65	94.73	81.93	91.96	80.16	74.45
  CenterPoint J6Gen2_Base/2.5.1	84.20	94.24	80.77	92.59	80.16	73.25

  Eval Range: 50.0 - 90.0m

  Model version	mAP	car (58,562)	truck (5,101)	bus (2,078)	bicycle (758)	pedestrian (10,283)
  CenterPoint J6Gen2_Base/2.6.1_quant	65.14	80.59	60.54	74.62	52.13	57.80
  CenterPoint J6Gen2_Base/2.6.1	65.60	81.03	61.17	75.08	52.45	58.27
  CenterPoint J6Gen2_Base/2.5.1	63.43	79.61	57.96	69.85	52.76	56.99

  Eval Range: 90.0 - 121.0m

  Model version	mAP	car (20,371)	truck (3,172)	bus (376)	bicycle (155)	pedestrian (2,794)
  CenterPoint J6Gen2_Base/2.6.1_quant	41.21	57.40	41.00	34.60	35.44	37.61
  CenterPoint J6Gen2_Base/2.6.1	41.52	58.37	42.35	35.37	34.32	37.19
  CenterPoint J6Gen2_Base/2.5.1	38.55	55.21	40.77	27.18	33.56	36.02

  Eval Range: 0.0 - 121.0m

  Model version	mAP	car (143,453)	truck (15,220)	bus (4,729)	bicycle (2,292)	pedestrian (32,498)
  CenterPoint J6Gen2_Base/2.6.1_quant	73.92	85.00	67.41	81.10	68.98	67.10
  CenterPoint J6Gen2_Base/2.6.1	74.26	85.33	68.00	81.44	69.07	67.45
  CenterPoint J6Gen2_Base/2.5.1	72.88	84.14	66.10	78.57	69.49	66.08

Latency (autoware centerpoint)

Observations

• INT8 significantly reduces average and median latency:
  • Mean latency improved by 25.1% (48.89 ms → 36.61 ms).
  • Median latency improved by 27.5% (47.99 ms → 34.78 ms).
• Throughput increased from 20.46 FPS to 27.31 FPS (+33.5%).
• Tail latency improvements are smaller:
  • P99 improved by 6.7%.
  • Maximum latency improved by 3.1%.
• INT8 latency distribution is wider than FP16, resulting in a higher standard deviation.
• Quantization effectively accelerates normal-case inference, while worst-case latency remains dominated by non-compute overhead.