<![CDATA[
<p>FPS = 1000 / 시간 (ms)</p><div class="table-wrap"><table data-ke-type="table" data-ke-align="alignLeft" style="width: 100%; height: 160px;" border="1" data-ke-style="style3"><tbody><tr style="height: 18px;"><td style="width: 7.48573%; height: 18px; text-align: center;"><span data-ke-size="size14">모델명</span></td><td style="width: 4.52413%; height: 18px; text-align: center;"><span data-ke-size="size14">Waymo mAP/L2</span></td><td style="width: 12.7259%; height: 18px; text-align: center;"><span data-ke-size="size14">컴퓨팅 자원 및 추론 속도</span></td><td style="width: 5.92687%; height: 18px; text-align: center;"><span data-ke-size="size14">T_inf (4090)</span></td><td style="width: 4.16141%; height: 18px; text-align: center;"><span data-ke-size="size14">Score</span></td><td style="width: 8.45083%; text-align: center;">Parameters</td><td style="width: 4.61211%; height: 18px; text-align: center;"><span data-ke-size="size14">GFLOPs</span></td><td style="width: 7.37945%; text-align: center;"><span data-ke-size="size14">3070 -> 4090 <br>T_inf</span></td></tr><tr style="height: 18px;"><td style="width: 7.48573%; height: 18px; text-align: center;"><span data-ke-size="size14">PV-RCNN++</span></td><td style="width: 4.52413%; height: 18px; text-align: center;"><span data-ke-size="size14">70.9</span></td><td style="width: 12.7259%; height: 18px; text-align: center;"><span data-ke-size="size14">TITAN RTX (1개) @ ~100ms</span></td><td style="width: 5.92687%; height: 18px; text-align: center;"><span data-ke-size="size14">~19.8ms</span></td><td style="width: 4.16141%; height: 18px; text-align: center;"><span data-ke-size="size14">70.9</span></td><td style="width: 8.45083%; text-align: center;"><span data-ke-size="size14"> 논문 명시 X <br>(구조상 PV-RCNN<br>보다 작음)</span></td><td style="width: 4.61211%; height: 18px; text-align: center;"><span data-ke-size="size14">166.6</span></td><td style="width: 7.37945%; text-align: center;"><span data-ke-size="size14"> </span></td></tr><tr style="height: 16px;"><td style="width: 7.48573%; height: 16px; text-align: center;"><span data-ke-size="size14">SECOND</span></td><td style="width: 4.52413%; height: 16px; text-align: center;"><span data-ke-size="size14">60.97</span></td><td style="width: 12.7259%; height: 16px; text-align: center;"><span data-ke-size="size14"><span style="color: #333333; background-color: #ffffff;"> GTX 1080 Ti (1개) @ ~</span>50ms</span></td><td style="width: 5.92687%; height: 16px; text-align: center;"><span data-ke-size="size14">~6.9ms</span></td><td style="width: 4.16141%; height: 16px; text-align: center;"><span data-ke-size="size14">60.97</span></td><td style="width: 8.45083%; text-align: center;"><span data-ke-size="size14">20M </span></td><td style="width: 4.61211%; height: 16px; text-align: center;"><span data-ke-size="size14">91.2</span></td><td style="width: 7.37945%; text-align: center;"><span data-ke-size="size14">~15.26ms</span></td></tr><tr style="height: 18px;"><td style="width: 7.48573%; height: 18px; text-align: center;"><span data-ke-size="size14">Part-A2-Anchor</span></td><td style="width: 4.52413%; height: 18px; text-align: center;"><span data-ke-size="size14">66.92</span></td><td style="width: 12.7259%; height: 18px; text-align: center;"><span data-ke-size="size14">NVIDIA Tesla V100 @ 71ms</span></td><td style="width: 5.92687%; height: 18px; text-align: center;"><span data-ke-size="size14">~12.15ms</span></td><td style="width: 4.16141%; height: 18px; text-align: center;"><span data-ke-size="size14">66.92</span></td><td style="width: 8.45083%; text-align: center;"><span data-ke-size="size14">244M </span></td><td style="width: 4.61211%; height: 18px; text-align: center;"></td><td style="width: 7.37945%; text-align: center;"><span data-ke-size="size14">~37.43ms</span></td></tr><tr style="height: 18px;"><td style="width: 7.48573%; height: 18px; text-align: center;"><span data-ke-size="size14">DSVT-Pillar</span></td><td style="width: 4.52413%; height: 18px; text-align: center;"><span data-ke-size="size14">73.2</span></td><td style="width: 12.7259%; height: 18px; text-align: center;"><span data-ke-size="size14">NVIDIA A100 (1개) @ 67ms</span></td><td style="width: 5.92687%; height: 18px; text-align: center;"><span data-ke-size="size14">~15.8ms</span></td><td style="width: 4.16141%; height: 18px; text-align: center;"><span data-ke-size="size14">73.2</span></td><td style="width: 8.45083%; text-align: center;"><span data-ke-size="size14">71M</span></td><td style="width: 4.61211%; height: 18px; text-align: center;"></td><td style="width: 7.37945%; text-align: center;"></td></tr><tr style="height: 18px;"><td style="width: 7.48573%; height: 18px; text-align: center;"><span data-ke-size="size14">DSVT-Voxel</span></td><td style="width: 4.52413%; height: 18px; text-align: center;"><span data-ke-size="size14">75.1</span></td><td style="width: 12.7259%; height: 18px; text-align: center;"><span data-ke-size="size14"><span style="color: #333333; background-color: #ffffff;">NVIDIA A100 (1개) @ </span>97ms</span></td><td style="width: 5.92687%; height: 18px; text-align: center;"><span data-ke-size="size14">~22.9ms</span></td><td style="width: 4.16141%; height: 18px; text-align: center;"><span data-ke-size="size14">~73.89</span></td><td style="width: 8.45083%; text-align: center;"><span data-ke-size="size14"> </span></td><td style="width: 4.61211%; height: 18px; text-align: center;"></td><td style="width: 7.37945%; text-align: center;"></td></tr><tr style="height: 18px;"><td style="width: 7.48573%; height: 18px; text-align: center;"><span data-ke-size="size14">PV-RCNN</span></td><td style="width: 4.52413%; height: 18px; text-align: center;"><span data-ke-size="size14">67.7</span></td><td style="width: 12.7259%; height: 18px; text-align: center;"><span data-ke-size="size14">TITAN RTX @ ~303ms</span></td><td style="width: 5.92687%; height: 18px; text-align: center;"><span data-ke-size="size14">~59.8ms</span></td><td style="width: 4.16141%; height: 18px; text-align: center;"><span data-ke-size="size14">~52.73</span></td><td style="width: 8.45083%; text-align: center;"><span data-ke-size="size14">50M </span></td><td style="width: 4.61211%; height: 18px; text-align: center;"></td><td style="width: 7.37945%; text-align: center;"><span data-ke-size="size14">~104.66ms</span></td></tr><tr style="height: 18px;"><td style="width: 7.48573%; height: 18px; text-align: center;"><span data-ke-size="size14">Voxel R-CNN</span></td><td style="width: 4.52413%; height: 18px; text-align: center;"><span data-ke-size="size14">65</span></td><td style="width: 12.7259%; height: 18px; text-align: center;"><span style="color: #333333; background-color: #ffffff;" data-ke-size="size14">RTX 2080 Ti @ ~40ms</span></td><td style="width: 5.92687%; height: 18px; text-align: center;"><span data-ke-size="size14">~6.5ms</span></td><td style="width: 4.16141%; height: 18px; text-align: center;"><span data-ke-size="size14">65</span></td><td style="width: 8.45083%; text-align: center;"><span data-ke-size="size14">28M</span></td><td style="width: 4.61211%; height: 18px; text-align: center;"></td><td style="width: 7.37945%; text-align: center;"></td></tr><tr style="height: 18px;"><td style="width: 7.48573%; height: 18px; text-align: center;"><span data-ke-size="size14">PointPillars</span></td><td style="width: 4.52413%; height: 18px; text-align: center;"><span data-ke-size="size14">62.8</span></td><td style="width: 12.7259%; height: 18px; text-align: center;"><span style="color: #333333; background-color: #ffffff;" data-ke-size="size14"> GTX 1080 Ti (1개) @ ~23.8ms</span></td><td style="width: 5.92687%; height: 18px; text-align: center;"><span data-ke-size="size14">~3.3ms</span></td><td style="width: 4.16141%; height: 18px; text-align: center;"><span data-ke-size="size14">62.8</span></td><td style="width: 8.45083%; text-align: center;"><span data-ke-size="size14">18M </span></td><td style="width: 4.61211%; height: 18px; text-align: center;"></td><td style="width: 7.37945%; text-align: center;"><span data-ke-size="size14">~16.60ms</span></td></tr><tr><td style="width: 7.48573%; text-align: center;"><span data-ke-size="size14">CenterPoint<br>-Pillar</span></td><td style="width: 4.52413%; text-align: center;"><span data-ke-size="size14">66</span></td><td style="width: 12.7259%; text-align: center;"><span style="color: #333333; background-color: #ffffff;" data-ke-size="size14"> <span style="color: #333333; background-color: #ffffff;" data-ke-size="size14">NVIDIA A100 @ ~35.7ms</span></span></td><td style="width: 5.92687%; text-align: center;"><span data-ke-size="size14">~8.4ms</span></td><td style="width: 4.16141%; text-align: center;"><span data-ke-size="size14">66</span></td><td style="width: 8.45083%; text-align: center;"><span data-ke-size="size14">5M </span></td><td style="width: 4.61211%; text-align: center;"><span data-ke-size="size14">493.11</span></td><td style="width: 7.37945%; text-align: center;"><span data-ke-size="size14"> </span></td></tr></tbody></table></div><p> </p><p>4090 환산 기준</p><p>GTX 1080 Ti에서 100ms 정도 걸렸다면 1080 Ti에서의 FLOPs를 확인해서 4090 기준으로 환산함</p><p> </p><p><a href="https://www.techpowerup.com/gpu-specs/geforce-gtx-1080-ti.c2877" target="_blank" class="ke-link">https://www.techpowerup.com/gpu-specs/geforce-gtx-1080-ti.c2877</a></p><div class="figure-img" data-ke-type="image" data-ke-style="alignLeft" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/1d75563b514759753813521c362872da1d50ab90" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/1d75563b514759753813521c362872da1d50ab90" data-origin-width="285" data-origin-height="259"></div><p> </p><p><a href="https://www.techpowerup.com/gpu-specs/geforce-rtx-4090.c3889" target="_blank" class="ke-link">https://www.techpowerup.com/gpu-specs/geforce-rtx-4090.c3889</a></p><div class="figure-img" data-ke-type="image" data-ke-style="alignLeft" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/2f4235ec7c7f4f22fbe19b09319676b184f071ef" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/2f4235ec7c7f4f22fbe19b09319676b184f071ef" data-origin-width="277" data-origin-height="258"></div><p> </p><p> </p><div class="figure-img" data-ke-type="image" data-ke-style="alignLeft" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/d9822ff9c9185d79acf8b5cd023872b52b022a82" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/d9822ff9c9185d79acf8b5cd023872b52b022a82" data-origin-width="199" data-origin-height="29"></div><p>1080 ti와 4090이 약 7.28배 정도 차이나므로 ms에 환산해서 보면 됨</p><div class="figure-img" data-ke-type="image" data-ke-style="alignLeft" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/78203ea4a559f17ea17cbf952ae57b833dc5c5c7" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/78203ea4a559f17ea17cbf952ae57b833dc5c5c7" data-origin-width="480" data-origin-height="23"></div><p> </p><p> </p><p>FLOPs 확인</p><div class="table-wrap"><table data-ke-type="table" data-ke-align="alignLeft" style="width: 100%;" border="1"><tbody><tr><td style="width: 23.8651%; text-align: center;">NVIDIA GeForce RTX 4090</td><td style="width: 13.3985%; text-align: center;">GTX 1080 Ti</td><td style="width: 17.1815%; text-align: center;">NVIDIA Tesla V100</td><td style="width: 13.2567%; text-align: center;">NVIDIA A100</td><td style="width: 15.1402%; text-align: center;">TITAN RTX</td><td style="width: 17.158%; text-align: center;">RTX 2080 Ti</td></tr><tr><td style="width: 23.8651%; text-align: center;"><span style="color: #333333;" data-ke-size="size16">82.58 TFLOPs</span></td><td style="width: 13.3985%; text-align: center;">11.34 TFLOPs</td><td style="width: 17.1815%; text-align: center;"><span style="color: #333333;" data-ke-size="size16">14.13 TFLOPs</span></td><td style="width: 13.2567%; text-align: center;">19.49 TFLOPs</td><td style="width: 15.1402%; text-align: center;"><span style="color: #333333;" data-ke-size="size16">16.31 TFLOPS</span></td><td style="width: 17.158%; text-align: center;"><span style="color: #333333;" data-ke-size="size16"> <span style="color: #333333;" data-ke-size="size16">13.45 TFLOPS</span> </span></td></tr></tbody></table></div><p> </p><p> </p><p> </p><p>pv-rcnn++</p><p><a href="https://arxiv.org/abs/2102.00463" target="_blank" class="ke-link">https://arxiv.org/abs/2102.00463</a></p><div class="figure-img" data-ke-type="image" data-ke-style="alignLeft" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/d857e25c2bf24252c8213b414e9b7d7b6bba4731" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/d857e25c2bf24252c8213b414e9b7d7b6bba4731" data-origin-width="495" data-origin-height="277"><div class="figcaption">PV-RCNN++ FPS</div></div><div class="figure-img" data-ke-type="image" data-ke-style="alignLeft" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/21b0b7e21e5a10e9ef02c10de8b769ea43300e4b" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/21b0b7e21e5a10e9ef02c10de8b769ea43300e4b" data-origin-width="441" data-origin-height="260"><div class="figcaption">1개 사용</div></div><p> </p><p>SECOND</p><div class="figure-img" data-ke-type="image" data-ke-style="alignLeft" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/ed3fdca5e06996313fd361b416ef367197a8c26d" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/ed3fdca5e06996313fd361b416ef367197a8c26d" data-origin-width="802" data-origin-height="320"></div><div class="figure-img" data-ke-type="image" data-ke-style="alignCenter" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/f45441e2865da7533d62ce0065a37e93ab19fcd2" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/f45441e2865da7533d62ce0065a37e93ab19fcd2" data-origin-width="770" data-origin-height="45"></div><p> </p><p> </p><p> </p><p> </p><p> </p><p>VoxelNeXt</p><p><a href="https://arxiv.org/pdf/2303.11301" target="_top" class="ke-link">https://arxiv.org/pdf/2303.11301</a></p><div class="figure-img" data-ke-type="image" data-ke-style="alignLeft" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/ddef11de84067dfc6e1c913d0f978e5e36124903" class="txc-image" width="700" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/ddef11de84067dfc6e1c913d0f978e5e36124903" data-origin-width="813" data-origin-height="481"><div class="figcaption">Waymo dataset</div></div><div class="figure-img" data-ke-type="image" data-ke-style="alignLeft" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/81883e1a6466253baa4e55ae84edf05c9136a5bc" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/81883e1a6466253baa4e55ae84edf05c9136a5bc" data-origin-width="491" data-origin-height="173"><div class="figcaption">추론 속도</div></div><p> </p><p> </p><p> </p><p>Part-A2-anchor</p><p><a href="https://arxiv.org/pdf/1907.03670" target="_blank" class="ke-link">https://arxiv.org/pdf/1907.03670</a></p><div class="figure-img" data-ke-type="image" data-ke-style="alignLeft" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/ed23f7917a08354718d7e0c186bc8632b31d9100" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/ed23f7917a08354718d7e0c186bc8632b31d9100" data-origin-width="465" data-origin-height="516"></div><p> </p><p> </p><p> </p><p> </p><p> </p><p>DSVT-(Pillar && Voxel)</p><p><a href="https://arxiv.org/pdf/2301.06051" target="_blank" class="ke-link">https://arxiv.org/pdf/2301.06051</a></p><div class="figure-img" data-ke-type="image" data-ke-style="alignLeft" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/4c6b4a6d203bcc0b2825030d8327dbf02d4f4b54" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/4c6b4a6d203bcc0b2825030d8327dbf02d4f4b54" data-origin-width="724" data-origin-height="458"></div><div class="figure-img" data-ke-type="image" data-ke-style="alignLeft" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/50eb6dc7469baf7fd453a7ea9fed2331b11ec4be" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/50eb6dc7469baf7fd453a7ea9fed2331b11ec4be" data-origin-width="534" data-origin-height="228"></div><div class="figure-img" data-ke-type="image" data-ke-style="alignLeft" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/5e38820d80f1d8c91e1678f1fb4e440592d7c30c" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/5e38820d80f1d8c91e1678f1fb4e440592d7c30c" data-origin-width="505" data-origin-height="143"></div><p> </p><p> </p><p> </p><p>PointPillar</p><div class="figure-img" data-ke-type="image" data-ke-style="alignLeft" data-ke-mobilestyle="widthOrigin"><img src="https://t1.daumcdn.net/cafeattach/1RgNt/2f052fc36586043bf0c37be0c94b5ac35122071a" class="txc-image" data-img-src="https://t1.daumcdn.net/cafeattach/1RgNt/2f052fc36586043bf0c37be0c94b5ac35122071a" data-origin-width="432" data-origin-height="224"><div class="figcaption">1개 사용</div></div>
<!-- -->
]]>
카페 게시글
과제게시판
[자율주행 AI 챌린지] 3D 후보 모델 선정
신민서
추천 0
조회 141
25.08.25 23:57
댓글 5
북마크 번역하기 공유하기 기능 더보기
다음검색
첫댓글 GPU 사양에 보면 초당 계산할수 있는 실수연산의 갯수(FLOPs)가 나와있음 이를 기준으로 환산하면됨
그대로 안되면 openpcdet 테스트코드를 직접실행해서 추론시간을 측정하고 4090기준으로 환산할것
너는 지난주에 도대체 뭘 한거냐?
정리 다했으면 openpcdet의 웨이모 데이터셋 테스트 코드를 이용해서 추론시간을 측정하고 계산치와 비교하여 검증할것
선행학습가중치가 필요할텐데 옵션에서 가중치 없이 랜덤설정해서 할수 있는지 확인하고 가중치가 반드시 있어야하면 키티가중치 파일을 사용할것 정확도는 필요없으니 가중치가 파일의 성능은 신경쓰지 말것
넵
테스트시 사용하는 입력데이터는 최소화하여 실행시간을 단축할것, 웨이모 데이터셋은 승현이가 가지고 있으니 다운로드할필요없음